Nati Shalom's Blog

Using SSD As A Foundation For New Generations Of Flash Databases

“You just can't have it all” is a phrase that most of us are accustomed to hearing and that many still believe to be true when discussing the speed, scale and cost of processing data. To reach high speed data processing, it is necessary to utilize more memory resources which increases cost. This occurs because price increases as memory, on average, tends to be more expensive than commodity disk drive. The idea of data systems being unable to reliably provide you with both memory and fast access—not to mention at the right cost—has long been debated, though the idea of such limitations was cemented by computer scientist, Eric Brewer, who introduced us to the CAP theorem.

The CAP Theorem and Limitations for Distributed Computer Systems

Through this theorem Brewer stated that it was impossible for any distributed computer system to be able to provide users with these three following guarantees simultaneously:

• Consistency (Every node will be able to view the same data at the same time)

• Availability (Every request will receive a response)
• Partition Tolerance (The system will continue to operate even if the system faces any arbitrary failures)

 We can evaluate different approaches for data management solution by the above three properties and the tradeoffs that each one of them faces when utilized over the other. For example, what is the tradeoff of putting more emphasis on consistency? The trade off will most likely be less availability or partition tolerance.

 Traditional RDBS solutions offered consistency over partition tolerance and cost. In-Memory caching solution such as memcache offered a different set of tradeoffs, for example speed over some degree of consistency (between the in-memory data and the data held in disk).In Memory Data Grid or In Memory Computing, as it is called today, extended the use of memory as the system of record and thus enabled higher degree of consistency by reducing the dependency on the underlying disk databases. The new generation of NoSQL databases took a different approach and offered speed and scale at low cost over consistency by utilizing distributed commodity storage and relying heavily on asynchronous data flow to speed up the data processing.

 Quite often there is a direct correlation between those tradeoffs and cost - for example high degree of consistency often relies on synchronous data-flow and replication operations which often come at a cost of speed. To overcome the speed limit we often need to use memory or other flavors of high speed storage which often translate to high cost. On the other hand if we can compromise on consistency, we can offer speed and scale using commodity resources (  without relying on high cost resources).

As technology advances and demands for cost effective and scalable, reliable data increase, new solutions are coming to light, with one of the most favored being a combination between two items we are already familiar with: RAM and flash.

 

Why SSD Is a Viable Storage Solution

Many of the underlying assumptions in the previously mentioned alternatives that  was built under the assumption that disk is the bottleneck. If disk is the bottleneck, we needed layers of optimizations to minimize the access to disk. SSD provides a high speed storage device, which gets rid of the idea that disk is the bottleneck. This is a fundamental change in one of core assumptions behind the design of many databases today, let me explain..

 For example if the disk is slow it makes sense to put various filters to minimize the access to disk. As disk gets faster all of those filters becomes an overhead. In many cases it would be faster writing directly to disk without those additional layers. In similar way the use of asynchronous operations need to change as well. When the disk is slow it makes sense to use asynchronous operations to deffer write operations. However, when the disk is no longer the bottleneck it would be faster and vastly simpler to access the disk directly. This also allows avoiding many of the consistency tradeoffs that i mentioned earlier that are often a result of those asynchronous operations.

 

Putting SSD and RAM Together using off Heap Storage

 RAM storage can be as big as necessary and it is incredibly quick—but it costs about ten times more than a flash disk. By putting SSD and RAM together we can optimize the cost/performance ratio. The main challenge in doing so, remain around consistency-i.e. how do we keep the data in-memory and flash in sync so that from an end user perspective this integration would look completely seamless?

One method of such integration is referred to as off-heap storage.  

Off Heap storage is often implemented as a plug-in that is used by the memory data-store to offload its data from the actual RAM into Flash and visa versa.

Traditionally Off Heap was implemented with shared memory which is basically a block storage that provides external access to the same RAM device. By doing so, we bypass the management overhead of managing data in RAM through JAVA. The limitation of that approach is that the capacity is still limited by the size of RAM. It also forces an external data management and garbage collection layer to manage this external heap.

With SSD, users can now have both the memory and the SSD device synchronized and regularly kept up-to-date, meaning storage is scalable and consistent.

Some of the SSD drivers provides a key,value interface which makes the integration vastly simpler as it takes care of its own data  management.  Having said that, most of the In-Memory implementation that uses SSD as off-heap storage provides fairly limited functionality in terms of query and transaction support. This is due to the fact that they( What is they?)  integrate with SSD through a disk storage interface.  

 

Using SSD as a Foundation for Flash DataBase

 To really make the best use of SSD and maximize its potential we need to think of SSD as foundation for database. This requires a more tight and native integration with SSD in order to overcome some of the limitations that exist with many of the current In Memory Data Grid and SSD implementations.

 The specific set of features that are needed for this sort of integration includes: 

• Portable and native Key/Value interface that works against any flash device.
• Using Flash as Durable Storage - Flash can be used as a persistent data-store and not just as extension to RAM. As such we can leverage the durability of flash to speed up the load time of data from the underlying flash device incase of a planned or un-planned recovery process.
• Support for batch operations - batch operations are a common way to speed up the access to flash by minimizing the number of cross boundaries calls between the RAM device and the underlying flash devices.
• Transactional - To support transactional access we need to extend the batch operations to fail, succeed as a single atomic operation.
• Query index support - to speed up the query and access time, the data needs to be indexed in a way that fetching a particular object by its IDcould point directly to its physical location on disk.

 

To make this a practical reality GigaSpaces partnered with SanDisk who through fusion I/O acquisition own the majority of the SSD market. As part of this partnership, we implemented a new version of our In Memory Data Grid, XAP MemoryXtend, which is now integrated with SanDisk ZetaScale interface. ZetaScale implements all of the above features as a general purpose flash disk interface. Through this integration we are now able to utilize the XAP RAM based storage for handling complex queries and analytics as well as transactional support. We can also leverage the XAP cluster support for integrating multiple flash devices that are mounted on various machines and make them work as one big transactional database ( as illustrated in the diagram below).

The XAP MemoryXtend cluster consists of a number of partitions each plugged into a local SSD drive.  Each partition can have at least one backup running on a separate machine to ensure  availability. The XAP client is a smart proxy that abstracts the underlying cluster and exposes all of the physical partitions through a single data-grid interface. The proxy routes the write or query requests to a particular partition in cases where we’re looking for a specific data item. In the case of agregated queries it will invoke a parallel query against all nodes and consolidate the results into a single result-set.

 

Flash DataBase as Service

As Flash devices are now being supported by cloud providers such as Amazon, it only make sense to leverage this capability and offer this sort of Flash based database on-demand.

This is where the XAP Deployment and Orechestration comes handy as it automates the process of deployment and management of XAP clusters across a variety of cloud infrastructures.

For more information, check out The next Big Disruption in Big Data.

The next big thing in big data: fast data

The big data movement was pretty much driven by the demand for scale in velocity, volume, and variety of data. Those three vectors led to the emergence of a new generation of distributed data management platforms such as Hadoop for batch processing and NoSQL databases for interactive data access. Both were inspired by their respective predecessors Google (Hadoop, BigTable) and Amazon (Dynamo DB).

As we move to fast data, there’s more emphasis on processing big data at speed. Getting speed without compromising on scale pushes the limits in the way most of the existing big data solutions work and drives new models and technologies for breaking the current speed boundaries. New advancements on the hardware infrastructure with new flash drives offer great potential for breaking the current speed limit which is bounded mostly by the performance of hard drive devices.

Why using existing RDBMS, NoSQL on top of flash drive is not enough

Many of the existing databases — including more modern solutions such NoSQL and NewSQL — were designed to utilize the standard hard drive devices (HDD). These databases were designed with the assumption that disk access is slow and therefore they use many algorithms such as Bloom filters to save access to disk in case the data doesn’t exist. Another common algorithm is to use asynchronous write to commit-log. A good insight into all the optimizations that are often involved in a single write process on NoSQL databases is provided through the Cassandra Architecture. Let’s have a look at what that entails below:

 

The Cassandra write path

 

What happens when disk speed is no longer the bottleneck?

When the speed of disk is no longer the bottleneck, as in the case of flash devices, then a lot of the optimization turns into overhead. In other words, with flash devices it will be faster and simpler to access the flash device directly for every write or read operation.

Flash is not just a fast disk

Most of the existing use cases use flash devices as faster disk. Using flash as a fast disk was a short path to bypass the disk performance overhead without the need to change much of the software and applications. Having said that, this approach inherits many unnecessary disk driver overhead. So to exploit the full speed potential of flash devices, it would be best to access the flash devices directly from the application and treat flash devices as a key/value store rather than a disk drive.

Why can’t we simply optimize the existing databases?

When we reach a point in which we need to change much of the existing assumptions and architecture of the existing databases to take advantage of new technology and devices such as flash, that’s a clear sign that local optimization isn’t going to cut it. This calls for new disruption.

The next big thing in big data

Given the background above, I believe that in the same way that demand for big data led to the birth of the current generation of data-management systems, the drive for fast data will also lead to new kinds of data management systems. Unlike the current set of databases, I believe that the next generation databases will be written natively to flash and will use direct access to flash rather than the regular disk-based access. In addition to that, those databases will include high performance event streaming capabilities as part of their core API to allow processing of the data as it comes in, and thus allowing real-time data processing.

 

Fast data in the cloud

Many of the existing databases weren’t designed to run in cloud as first-class citizens and quite often require fairly complex setup to run well in a cloud environment.

As cloud infrastructure matures, we now have more options to run big data workload on the cloud. The next generation databases need to be designed to run as a service from the get-go.

To avoid the latency that is often associated with such a setup, the next generation databases will need to run as close as possible to the application. Assuming that many of the applications will run in one cloud or another, it means that those databases will need to have built-in support for different cloud environments. In addition, they will need to leverage dynamic code shipping to pass code with the data, and in this way allow complex processing with minimum network hops.

In-memory databases and data grid are the closest candidates to drive next-gen flash databases

RAM and flash-based devices have much more in common than flash and hard drive. In both RAM and flash, access time is fairly low and wasn’t really the bottleneck; in-memory databases use direct access to the RAM device through key value interface to store and index data.

Those factors makes in-memory based databases more likely to fit into the next generation flash databases.

The combination of in-memory databases and data grid on top of flash devices also allows the system to overcome some of the key capacity and cost per GB limitations of an in-memory based solution. The integration of the two will allow an increase in the capacity per single node to the limit of the underlying flash device rather than to the limit of the size of RAM.

The overall architecture of the integrated solution looks as follows (Source: Gartner):

 

 

As can be seen in the diagram above, the IMC (in-memory computing) layer acts as the front end to the flash device and handles the transactional data access and stream processing, while the flash device is used as an extension of the RAM device from the application perspective.

There are basically two modes of integration between the RAM and flash device.

LRU Mode – In this mode, we use the RAM as a caching layer to the flash device. The RAM device holds the “Hot,” i.e. the most recently used, and the flash device holds the entire set.

Pros: Optimized for maximum capacity.

Cons: Limited query to simple key/value access

Fast Index Mode – In this mode, all the indices in RAM and the data in flash holds the entire set.

Pros: Supports complex querying, including range queries and aggregated functions.

Cons: Capacity is limited to the size of indexes that can be held in RAM.

In both cases, the access to flash is done directly using key/value interface and not through disk drive interface. As both RAM and flash drives are fairly fast, it is simpler to write data to the flash drive synchronously and in this way avoid potential complexity due to inconsistency.

It is also quite common to sync the data from the IMC and flash devices into an external data store that runs on a traditional hard drive. This process is done asynchronously and in batches to minimize the performance overhead.

Direct flash access in real numbers

To put some real numbers behind these statements, I wanted to refer to one of the recent benchmarks that was done using an in-memory data grid based on GigaSpaces XAP and direct flash access using a key/value software API that allows direct access to various flash devices. The benchmark was conducted on several devices as well as on private and public cloud-based services on AWS. The benchmark shows a ten times increase in managed data without compromising on performance.

 

 

 

For full details of the benchmark, see the following white paper:

XAP MemoryXtend -Massive Application Storage Capacity for Real-Time Applications

 

 

 

Notes from Big Data Business Challenges Panel Discussion

Most of the Big Data topics that I covered in the past were focused mostly on the technical aspects of running Big Data.

In this post, I want to take a different route and discuss the business challenges based on a panel discussion that I had the pleasure of moderating.

 

 

Is Big Data a Big Hype?

One of the first questions that came up in the discussion is whether Big Data is over-hyped due to over-use in marketing content. Many of those who claim to provide either a Big Data solution or implement Big Data are just using the term to re-brand their product or offering.

 There was consensus across the panel that while Big Data is probably over hyped like any big trend, it does represent a real change in both technology and business. Let me explain:

In a pre-Big Data world, Big Data lived in expensive and closed data warehouses, where many of the solutions were often tied to specific, high end hardware. In addition, the term “warehouse” in itself meant that most of the data had little exposure and impact on the business.  Today, Big Data is everywhere and can be done at a fraction of the cost of the traditional data warehouses. With the addition of cloud the barrier to running Big Data has become significantly lower and now every startup can run Big Data almost at the scale of Google.

Having said that, the actual adoption of Big Data tends to vary quite significantly per industry with telecommunications (67%), consumer goods (57%) and financial services (52%) industries as the leaders in recognizing that Big Data can greatly improve their understanding of customer needs according to a recent Hitachi Data Systems Corporation (HDS) study  .The same study indicates that more than 60% of the firms in the financial services and consumer goods industries haven't started any Big Data programs. Healthcare and life science are lagging further behind; 72% of them haven't started any Big Data programs. This means that the perspective of whether Big Data is real or just marketing hype tends to vary as well depending on the industry.

 

How do you make sense out of your Big Data? Do we need a new role for Chief Data Officer?

Many of the questions in the panel had to do with how to make sense of the data.

One of the common strategies to make sense of the data is to assign a data scientist or Chief  Data Officer, expecting them to deal with the data and offer useful insights. Another common strategy is to put all the data in something like Hadoop with the hope that having the data in hand will provide a variety of interesting insights that will change the way we operate.

Putting the data in Hadoop or assigning a specialized data analyst are often just the beginning of a process.

Eyal Veitzman from Wix described how they deal with the challenge. To really make sense of your data, you first need to change the organizational structure in which the business, developers and data analysts will be part of the same group. In this way you can turn insights into action and more importantly have a faster feedback loop that will allow you to iterate quickly.  

 

What is the business value behind Big Data?

The business value behind Big Data comes in different shapes and forms depending on where you apply it in your system. Insightera uses Big Data to capture user interaction and increase the conversion of users using personalized responses. Wix uses Big Data to optimize their user conversion and satisfaction.  Nice Systems uses Big Data to improve their call center user satisfaction. PayPal is using Big Data in all sort of shapes and forms to both detect fraud as well as to optimize their operations. Check Point uses Big Data to collect data from all their gateways and in that way keep their gateways protected from new virus attacks

So how much can we gain from applying Big Data practices in our organization?

According to the Hitachi Data Systems Corporation (HDS) study, Big Data can help companies' revenue by 25%. This applies mostly to companies that use Big Data to optimize their internal operations and products, like most of the companies in the panel. That is obviously more true for companies that sell Big Data insights as part of their business, as in the case of Insightera.

Is there a good visualization tool for Big Data?

Big Data visualization is a tough challenge and there is no simple answer or single framework that fits them all. Using infographic style can help to visualize data in a way that makes it easier to capture lots of data into a single, readable report using a variety of visualization dimensions, such as fonts, colors, graphics, etc. There are many open source frameworks, such as Jaspersoft and many SaaS-based frameworks, aimed to provide specialized reporting and visualization tools. A useful reference on that regard can be found here here: The 36 best tools for data visualization.

Having said that, the important thing is to decide what and how to measure. Quite often, the number of Key Performance Indicators (KPIs) must be limited to a fairly small number.

The second important thing is to realize that as humans, we are not designed to capture and process vast amounts of information and therefore, visualization tools provide only a small aid to a much bigger challenge.

In summary, there is a dangerous slippery slope with Big Data visualization; fairly quickly we can flood our users with many graphs and numbers to a point where what we actually get is a dimension return and more complexity and confusion than clarity. Those who had experience with analytics tools such as Google Analytics are probably familiar with that experience. On the other hand, narrowing down the key performance indicators into a small number of KPIs makes the visualization challenge simple.

 

Moving from Reports to Actions

The key in capturing most of the value of Big Data is to turn Big Data into actions. In the example of Insightera, actions means presenting a user with the most relevant information on a website. In the case of Nice Systems, actions would refer to routing the caller in a call center to the right agent based on the caller’s profile and history of behaviour. In the case of GigaSpaces, this would mean constantly adjusting our provisioning system to detect failures and auto-heal a production environment when something goes wrong.

To get to that level we need to apply automation practices on how we collect, analyze and execute actions out of our data. We also need to be able to process data in real time to reduce the time to action.

Final words: How do make our data instantly actionable?

It is quite clear that the end game of any Big Data initiative is to figure out a way to make it instantly actionable. One of the big challenges on that regard is that quite often we need to compromise between scale and speed. In fact, many of the Big Data systems (Hadoop included) use batch processing to be able to process large amounts information, meaning that the actual insights, and thus actions, from that data will be available only when the batch cycle has been completed.

To make Big Data instantly actionable we need to figure out a way in which we can process vast amounts of data in real time and still keep the cost down.

During the panel discussion, PayPal described their use of In-Memory based architecture to handle this challenge. Still, pure In-Memory-based solutions tend to be fairly expensive and limited in capacity comparing to disk based alternatives. So the next big thing IMO is to devise a way in which we can get In-Memory speed at a cost that is closer to Disk. This is where new advancements in SSD have the potential to disrupt the entire way in which we manage and process Big Data, by providing high capacity durable RAM based out of SSD that can reach terabytes capacity per single machine. The fact that it can be durable means that we could rely completely on In-Memory databases to manage and store larger chunks of our Big Data with no need to offload or reload the data to and from external disks. I also expect that this will lead to the rise of a new generation of In-Memory databases that will be designed primarily for this new kind of durable RAM devices.

 

About the Panel

The Big Data Business challenges panel  was led by the CICC and in collaboration with IGTCloud, focusing primarily on the business challenges side of Big Data. The event was held simultaneously in California and Israel and was hosted by eBay. The panel included a range of industry experts in Big Data from a wide range of companies and industries and included Allen Kamer, Chief Commercial Officer, Optum Analytic (Co-founder of Humedica that was acquired by Optum Health), Doron Simon, Head of Cloud, Business Development, Nice Systems, Mickey Alon, Co-Founder & CEO, Insightera (Acquired by Marketo), Dr. Nachum Shacham, Principal Data Scientist, PayPal, Eyal Veitzman, Director of Business Operations, Wix and Ron Davidson, Head of Security Services, Check Point.

I had the special honor to organize and moderate the panel together with Ido Sarig, General Manager, IoT Solutions Group while Wind River moderated the panel in California.

 

Useful References

• Big data can help companies' revenue by 25%

Telecommunications (67%), consumer goods (57%) and financial services (52%) industries are leaders in recognizing that big data can greatly improve their understanding of customer needs. However, more than 60% of the firms in the financial services and consumer goods industries haven't started any big data programs. Healthcare and life science are lagging further behind; 72% of them haven't started any big data programs. http://timesofindia.indiatimes....­

• Fast Data hits the Big Data fast lane

The usual factors: the same data explosion that created the urgency for Big Data is also generating demand for making the data instantly actionable. Bandwidth, commodity hardware and, of course, declining memory prices, are further forcing the issue: Fast Data is no longer limited to specialized, premium use cases for enterprises with infinite budgets. Not surprisingly, pure in-memory databases are now going mainstream: ..

..Potential use cases for Fast Data could encompass:

• A homeland security agency monitoring the borders requiring the ability to parse, decipher, and act on complex occurrences in real time to prevent suspicious people from entering the country
• Capital markets trading firms requiring real-time analytics and sophisticated event processing to conduct algorithmic or high-frequency trades
• Entities managing smart infrastructure which must digest torrents of sensory data to make real-time decisions that optimize use of transportation or public utility infrastructure
• B2B consumer products firms monitoring social networks may require real-time response to understand sudden swings in customer sentiment
• The 36 best tools for data visualization

A useful reference for popular visualisation frameworks 

Industry Survey Shows: Organizations Turn to Stream Processing for their Big Data Needs

In a stream processing model, data is processed as it arrives to the Big Data system. With a batch processing model, only once the data is stored does the system run a variety of batch analytics - typically through a map/reduce style of processing.

 Facebook and Twitter are good examples of stream processing models in use today, as I outlined in a previous post entitled Facebook’s vs Twitter’s Approach to Real-Time Analytics.

 According to a recent survey of approximately 250 respondents, there is a trend moving towards stream processing in order to speed up analytics, with a significant increase in the popularity of this model when compared with last year. The number of organizations planning to use stream processing in 2014 has more than doubled (24%) from last year’s amount (10%).

This goes hand in hand with the fact that real-time analytics is also becoming more mainstream, as I pointed out in my 2014 Predictions.

Another interesting data point from the survey is that roughly 43% of the respondents defined real-time as sub-second and 42% defined it as sub-minute. This difference is quite interesting and could lead to different approaches for implementing stream processing.

For example, real-time is defined differently by Facebook and Twitter. In the case of Twitter, they chose Storm as their event processing engine, allowing them to process events at a sub-second resolution. Meanwhile, Facebook defines real-time at a 30 second batch window, choosing a logging-based approach to fit their need.

Based on the survey, it appears that both approaches are valid and can be applied in correlation to the degree of real-timeliness of your analysis.

 

Facebook Log-Centric Stream Processing	Twitter Event-Centric Stream Processing

In-Memory Data Grids become a more popular choice for real-time processing

According to the survey, 64% of the respondents indicated that they plan to combine In-Memory-based solutions for delivering their real-time analytics processing. This is also consistent with last year survey done by Ventana Research.

 I believe that will see an even bigger movement in this direction, as the cost/performance ratio of In-Memory Data Grids and In-Memory Databases would go down significantly with the combination of RAM and Flash devices, which together provide a fairly compelling solution from both performance and cost ratio to that of other disk-based alternatives.

 

Big Data in the Cloud increased to 56%

New developments in cloud infrastructure, such as the support for bare metal in the new OpenStack releases, as well as support, high memory, flash disk, etc, are removing many of the technical barriers for running I/O intensive workloads like Big Data in the cloud.

Indeed, according to the survey, there is a significant shift towards the use of Big Data in the cloud with 55% of organizations either using or planning to run their Big Data in the cloud in 2014.

 

Where do we go from here? How does this effect GigaSpaces’ future roadmap?

There are multiple areas in the GigaSpaces roadmap that aim to address this demand. 

• Real-Time Processing through Storm Integration - In this project, we integrate Storm on top of a Memory backend for both stream processing and data-store.
• Support for Flash Disk - The integration with flash disk is planned for our upcoming XAP release and will include support for SanDisk, Fusion I/O and other flash disk devices.
• Big Data in the Cloud - through Cloudify - We've been continuously extending our Big Data portfolio support and recently added support for Storm and Cognos, adding to our existing support for Hadoop, MongoDB, Cassandra, ElasticSearch, etc.

 

References

• GigaSpaces Annual Big Data & Cloud Survey 2013
• Eight Cloud and Big Data Predictions for 2014
• GigaSpaces and Storm Part 2: XAP Integration
• GigaSpaces And Storm Part 1: Storm Clouds
• Big-Data-Real-Time-Performance - Enjoying both worlds with one Architecture!
• 2012 Big Data Survey Report
• In Memory Computing (Data Grid) for Big Data

Big Data on OpenStack

Big Data systems by their very nature tend to be...Big.

Big in the amount of data, in the number and size of infrastructure behind it, etc. Cloud based infrastructure can be a good fit as a cost effecive instrastrucutre for running those Big Data systems. While this may sound obvious, many of the Big Data deployments that I've encountered run outside of a cloud environment. The main reason is the performance overhead that is often associated in running on a virtualized cloud environment. Let me explain..

The Performance Overhead of Running Big Data on the Cloud

A recent benchmark done by Petersenna over various virtualization solutions, such as XEN, VMWare, HyperV etc, shows that on average the performance overhead associated with virtualization can lead to 2.4 times slower disk latency and 25% slower network I/O.

How does the performance overhead translate into cost?

Based on the various benchmarks, I think that it would be fair to assume that running I/O intrusive workloads such as Big Data on a virtualized infrastructure would require 3X more resources than its Baremetal equivalent.

As we are talking about Big Data infrastructure - 3x means a lots of wasted resources. The operational costs coupled with the complexity associated with running a significantly bigger system yields a fairly substantial overhead and cost. In addition, the Network and Disk overhead are not deterministic and can vary quite substantially when the utilization gets higher. This leads to not just performance overhead, but also non-deterministic overhead. In Big Data terms, that means that a query for particular data can take 10msec one time or 30msec another time. Quite often, running Big Data analysis requires a sequence of those operations. Therefore, if we pile this overhead in a sequence of 10 queries, the variance in the response time can vary quite substantially.

While the choice of running Big Data on the cloud holds a lot of promise, the performance and non-deterministic behavior makes that choice limited to more of a niche scenario where this overhead  becomes less significant in comparison to the elasticity benefit of the cloud. For example, for sporadic worklads in which we run our analysis for a certain period of time and then can release the resources, using on-demand infrastructure is still a better choice, as running 3x the amount of resources for an hour is significantly cheaper than having a third of that infrastructure allocated 24/7.

OpenStack Bare Metal Cloud

The analysis above shows that for I/O intensive workloads, virtualized infrastructure isn't such a good fit.

Cloud is often viewed as an infrastructure on top of virtualization. Therefore, by definition, a cloud-based infrastructure inherits the benefits and limitations of virtualization.

Is the coupling between cloud and virtualization really mandatory?

If we think of a cloud as an infrastructure for getting compute, storage and network resources on-demand, then there is nothing  to necessitate the coupling of the cloud with virtualization. It became common to pair the two mainly because of the complexity involved with provisioning non-virtualized resources and the limitations in enabling partitioning of a given bare metal machine.

As I noted in one of my previous posts Bare Metal Cloud/PaaS,  there are ways today to provision a baremetal machine from an image and partition it just like we would do with a hypervisor based VM.  There are already cloud providers that offer a choice of bare metal machines as part of their cloud infrastructure.

A new bare metal project in the OpenStack Grizzly release takes this a step further. It allows us to use the same compute API (NOVA) and allocate a bare metal machine just as we would with a virtualized machine.

All we need to do to make the switch is to change the image type of our target machine and the cloud infrastructure will know to map that request and allocate a bare metal image instead of a virtualized instance.

With this option we can now run our entire Big Data workload on the cloud and not worry about switching environments depending on the workload happen to be sporadic or I/O intesive.

HubSpot OpenStack – Bare Metal Case Study

During the last OpenStack Summit, Jim O’Neill CIO at HubSpot showed how using OpenStack with a combination of a public virtualized cloud and private bare metal cloud enabled a 4X increase in their infrastructure efficiency:

 

“We took this single image, picked it up from public cloud into a Rackspace-powered private cloud and saw a 4X increased efficiency running that workload.”

 

Moving from Existing Data Centers to the Cloud

Many existing Big Data and BI systems run on traditional data center environments. Moving those systems into an OpenStack-based environment isn't going to be a walk in the park. This is where automation frameworks, such as Cloudify with the combination of Chef and Puppet, make this transition smoother.

In this approach, we can automate the deployment of our existing Big Data/BI systems in a way that will be abstracted from the underlying infrastructure. We can later use this abstraction to run Big Data systems in our existing data center, and when ready, we can use the same deployment framework on an OpenStack-based environment without re-doing any of that investment.

 

 

Learn more through hands-on experience - a Real Life Experience on HP OpenStack

As it often happen in this sort of conceptial discussions the points and arguments from this post may often sound artificial and not easy to grasp. To make it more down to earth, we've put a NoSQL datastore such as Couchbase, MongoDB, Cassandra, ElasticSearch and Big Data applications available on-demand on HP OpenStack cloud services. You can use this reference to launch any of those applications and then use the management console to browse through the recipes and customize it to your environment as needed.

It is also worth pointing out that the recipes behind this project are available on Github and can be easily deployed on your cloud or in your data center or even desktop in pretty much the same way. If you are interested in more details, please post a comment on this post or in the Cloudify forum.

For more on Big Data on OpenStack, come check out my presentation at Cloud Expo East on Tuesday, June 11th at 8:15am in the “Cloud Computing and Big Data” Track.

References

• Bare Metal Cloud/PaaS
• OpenStack Baremetal Project
• Big Data in the Cloud
• Big Data in the Cloud using Cloudify
• Putting Hadoop On Any Cloud (A video presentation)
• In Memory Computing (Data Grid) for Big Data
• Using the Cloudify Player as an Open Source Framework for Building Your Own Cloud Application Marketplace on OpenStack
• Going native: The move to bare-metal cloud services
• New bare metal cloud offerings emerging
• How much overhead does x86/x64 virtualization have?
• Amazon EC2 versus Bare Metal and KVM? The inside story on what you thought you knew about EC2

Java and Center Stage: Meet us at JavaOne 2011

Next week, I’m going to be in JavaOne. Unlike last year, where I expressed lots of skepticism about the way Oracle drives the Java Community, it seems that this year things are starting to get back on course – not that the skepticism has vanished completely, but at least there is a stronger sentiment that things are starting to settle down and take a more positive course.

It seems that now Java is taking center stage in the Cloud world, with more application platforms such as SalesForce/Heroku, VMware, Redhat/OpenShift, Cloud.com, JClouds, and obviously GigaSpaces providing a rich set of offerings that are based on Java.

One of the most interesting things in my opinion is that the combination of Cloud and Java brings new opportunities that until recently were unheard of:

 

• Heroku, which was very much Ruby-centric, added support for Java.
• GigaSpaces recently announced our partnership with Microsoft to bring Java to Microsoft's Azure community.

What’s interesting in those two examples is that the Cloud forces us to break away from the language wall. It brings communities that were living in completely isolated silos much closer together.

This forces a very different way of thinking about Java - and about the next generation Java platform specifically.

The new platform should be:

• Open for other languages (through integration points like Protobuff, Thrift, REST, and more, as well as the ability to run those languages)
• Embrace NoSQL implementations as an integral part of the platform
• Add support for dynamic data structures such as JSON, and schema-free document models
• Bring Dev and Ops closer by providing better exposure of operational aspect as part of the development frameowork
• Make cloud deployment a first class citizen

It seems that Java EE 7 is taking the right path in that direction. From Java EE 7 and Cloud Computing:

• The Java EE platform architecture is taking into account the operational impact of the cloud, more specifically by defining new roles, such as a PaaS administrator.
• The Java EE platform may also establish a set of constraints for PaaS-specific features, such as multi-tenancy, that deployments may have to obey. Applications may also be able to identify themselves as designed for cloud environments.
• All resource manager-related APIs, such as JPA, JDBC and JMS, will be updated to enable multi-tenancy. The programming model may be refined as well by introducing connectionless versions of the major APIs.
• Java EE will define a descriptor for application metadata to enable developers to describe certain characteristics of their applications that are essential for the purpose of running them in a PaaS environment. These may include being multitenancy-enabled, resources sharing, quality of service information, dependencies between applications.

What This Means

Java is regaining some of its luster from the last few years. It was being perceived as vaguely pedestrian and ordinary, and languages like Ruby and Python were being seen as more dynamic environments for rapid and scalable deployments.

Now we're seeing Java as the lingua franca of enterprise development. With so many environments being heterogeneous, a homogenous platform becomes a very desirable resource, and the JVM is able to run many languages in an integrated environment: Java, Ruby, Python, Scala, and Groovy, to name only a few, and many of these can run as compiled code and as scripts.

Even beyond the multiple programming language support of the JVM, you have very broad support for almost any number of remote procedure call mechanism you can think of. The old J2EE approach, where you defined your requirement and there was a specific way to fulfill it (which meant that every architecture was more or less nudged to the middle), is no longer a real limit.

If you can do it, you can do it in Java, without Java forcing you to sacrifice in the process.

This is a very powerful concept.

It puts your organization back under your control. This is desirable because you know your product and architecture better than a community process does. You know whether you need to support REST here, and SOAP there; you know you need to use NoSQL in this area and MySQL in that area.

A Remaining Challenge

Java EE 7 may be becoming aware of the new cloud-based environment, but it's still only one aspect of application deployment and design. What it does not address is the entire application environment.

This is where applications such as Cloudify come in. Cloudify treats every artifact involved with your deployment as a managed element.

Consider: when you deploy an enterprise application, you're not just deploying some operational code. That code has external requirements, like a database (or NoSQL warehouse); it might also rely on an Apache-based load balancer, or perhaps another independently-deployed artifact.

That's quite a burden on operations, because all of these are separately deployed and managed.

Cloudify, however, centralizes the management of each element your application is composed of, over the entire application's lifecycle, and provides a bridge to whatever environment you wish to use.

For example, you could have a MySQL database, used by a Java application hosted by a single Tomcat instance, fronted by an Apache httpd server, hosted on an internal Linux server. Cloudify can deploy and monitor each of those elements: the MySQL database, the schema and initial data; Tomcat; the Java application; Apache httpd, and the server upon which it all runs.

Now let's look at a larger deployment: a set of MySQL servers, a cluster of GigaSpaces XAP nodes, an instance of GlassFish hosting some web services, six .Net-based web services, and five httpd servers being load-balanced, deployed on a platform such as Microsoft Azure. From Cloudify's perspective, the deployment and management process is the same. You have more artifacts to deploy, and you're telling Cloudify to deploy on a different platform... that's it.

Cloudify can make sure each resource is running optimally, as well. If you can measure an attribute - such as CPU load, or disk space, or free memory - you can tell Cloudify to assert that some headroom exists, and what to do if a metric is exceeded.

For example, if you have a CPU running at 100% for an extended period of time, Cloudify can automatically deploy another instance of the component consuming the CPU, and configure load balancing so your application is more scalable, without human intervention.

Likewise, if you have four servers running a given process, all at 0% CPU load, you can reduce the number of containers with that component so that you're not overallocating resources, saving energy and money.

The key thought here:

You know how to solve your problem. We can help.

We'll be on the floor at JavaOne 2011, in booth #5002. We'll be happy to show you both XAP, our industry-leading application platform, and Cloudify, our deployment solution for the cloud-enabled world.

 

 

(This post was co-written with Joseph Ottinger, who unfortunately won't be able to make it to JavaOne this year. But he says that the best people in GigaSpaces will be there, so you're not missing out!)

Multi-tenancy: does it have to be that hard?

Multi-tenancy is a term that is used often in the context of SaaS applications. In general, multi-tenancy refers to the ability to run multiple users of an application on a shared infrastructure. The main motivation for doing this is efficiency, or in other words -- reducing the cost per user in comparison to a dedicated system where each user has their own dedicated environment.

Over the past few months I've had discussions with various SaaS providers, in which I tried to learn about the main challenges that SaaS providers face when they deal with multi-tenancy. As it happens, I came across an interesting article on this very topic, Why Multi-tenancy Matters in the Cloud by Alok Misra. The article later sparked an interesting discussion thread on the cloud mailing list, which I also find to be a good reference for diverse thoughts on this subject.

In this post I want to summarize some of my findings on multi-tenancy, as well as suggest a model that can make multi-tenancy significantly simpler.

The common challenge: Multi-tenancy at the data layer

There are many approaches and levels of multi-tenancy, depending on the application layer and type of application you're dealing with. For example, trying to separate user accounts in a CRM application is quite different than separating various batch processing jobs in a risk management application. That said, what seems a common challenge in almost every SaaS application is multi-tenancy at the data layer. To put it simply, the challenge is how to share data resources (database in most cases) among multiple users, while at the same time ensuring data isolation between those users, as if they are running on completely physically seperated servers.

In the remaining part of this post I'll address how multi-tenancy is currently being dealt with at the data layer, and how the existing challenges can be met.

How is multi-tenancy being done today?

For this part of the story, I think it might be useful to first look at the two extreme approaches: The fine-grained approach and the cross-grained approach.

Fine-grain multi-tenancy (Salesforce.com model)

For many SaaS providers the “poster child” of multi-tenancy is Salesforce.com. There is a very useful videocast explaining the Salesforce approach to multi-tenancy. In a nutshell, their approach is to share the same database instance between different users, and define the data model so that each table has a primary key identifying the customer ID. In addition, there is a separate metadata table that translates every query into a user-specific query before the query hits the database.

• Pros:
  • Very fine-grained multi-tenancy resulting in optimum cost efficiency per user.
• Cons
  • Complexity: Requires changes in the data model.
  • Doesn’t fit to existing application: Existing applications require a complete rewrite.
  • Modest degree of isolation: Users sharing the same resource can impact one another under heavy load situations. A maintenance/patch release required for a particular user might affect other users. In a more extreme scenario, there is also exposure to a higher degree of vulnerability. For example, a bug in the query mapping could point one user to data belonging to other users just by picking up the wrong customer ID.

Cross-grain multi-tenancy (hosted service model)

Most hosting providers (including hosted IT providers) enable outsourcing of certain IT applications. Efficiency is gained by sharing maintenance overhead, labor costs, and other data center costs such as network, electricity, and more.

• Pros:
  • No code changes are required, resulting in greater simplicity.
  • High degree of isolation.
• Cons
  • Lower efficiency and therefore limited cost margins.

Challenges with today's approach to multi-tenancy

The challenge with the fine-grained model is that it is fairly complex to maintain and implement. If you have an existing application, it requires a complete rewrite and also forces fairly significant changes in your existing data model. The challenge with the cross-grained model is that cost margins per user are still rather high. In a competitive environment, even a single competitor offering similar servicse to yours using the fine-grained approach could easily put you out of business.

The elasticity challenge

The other challenge that neither approaches addresses today is dynamic elasticity. What happens when a customer grows beyond their allocated shared capacity? Both approaches seem to rely on certain capacity assumptions, and can scale by splitting multiple users between their shared resources. However, none of the approaches seem to handle a situation where a particular user's capacity grows beyond the allocated resources.

Jayarama Shenoy provides a good introduction to that challenge in his comment in the discussion thread:

So, how does elasticity work in a multi-tenancy case? Say 100 tenants share an app on a server (by which I would think they go all the way down to a shared database). Then,say, five of them grow quite rapidly such that they need to be moved off that server, and the other 95 can continue to exist on the original server.
So 'something' in the application has to know how to pick off the records that belong to these five, copy it off into a second instance of that database and kill, move & restore application service for those tenants. (Of course, without being privy to the data contained in those records). And of course there's an implied load balancer element in front of all this to keep up appearances to the clients.

Based on this analysis I believe that the main challenge with the current approaches is that we have too few options available. No option seems to provide a good solution to the elasticity challenge, and this is what is making multi-tenancy “so damn hard”. This is especially true for existing businesses with existing products.

The solution: Making multitenancy simple

Learning from the past operating system experience

Some of us probably remember the days when Bill Gates came up with DOS, which was basically a single-user system. At the time, if you wanted to provide a multi-user application you had to write it yourself, and sometimes that required going down to the core interrupt level and program in assembly language. This made multi-user systems on DOS very uncommon. Unix, on the hand, was built for multitasking, multi-user environments, and therefore it was quite common to build multi-user systems in the Unix environment. Later on, when Windows NT came out, all that changed. Today we can’t even imagine a scenario where we would build an application that wouldn’t support multiple users. 

Lesson learned:

As long as the underlying infrastructure (the operating system in the example, and the database in our specific discussion) is not built for multi-tenancy, trying to solve the multi-tenancy challenge at the application level is doomed to be as complex as writing a multi-user system on DOS. 

In fact, once multi-tenancy was dealt with properly at the operating system level, writing a multi-tenant application became significantly simpler and therefore widely adopted. The same should apply to databases as well.

The database is the problem!

Almost all of the attempts I've seen in both the Salesforce approach and the hosted environment approach were built under the assumption that the database is a fixed “concrete” construct that cannnot be broken down easily into “smaller bricks”, let alone scale dynamically. It is therefore not surprising that most of the attempts to deal with multi-tenancy have tried to address the challenge outside the database.

Now, imagine what your approach to the challenge would be if a database could be easily broken down into small “bricks” without huge overhead, and if it could scale dynamically?

Multi-tenant data service

Similar to the experience we had with operating system, I believe that the solution for multi-tenancy has to be dealt with at the infrastructure level. In our specific discussion, that would be the database. In this section I’d like to draw the principles for a multi-tenant data service:

• Simplicity: With a multi-tenant database, applications can be written to a single database tenant just as they are today. How tenants are allocated and shared with other tenant needs would be completely abstracted from the application.
• Efficiency: Efficiency is maintained by the fact that all data tenants can share the same hardware (memory, CPU) with other tenants.
• Dynamic scaling: Each database tenant can be distributed and scale dynamically across multiple machines, if necessary, to meet demand. This includes moving data to other machines if necessary. Scaling must be supported seamlessly without any downtime.
• Support for multiple isolation/sharing levels: With a multi-tenant database, the level of sharing/isolation is set on demand, based on the specific requirements. At that level, a user can choose a dedicated tenant, in which case the multi-tenant database would allocate tenants on a set of machines that is dedicated to that specific user. Clearly, the benefit of this is isolation, at the cost of efficiency. Another type of sharing is available to the user at the same time: Choosing to share with any individual or group under the user's control. In other words, rather then dictating either extreme (full isolation or full sharing) for all users, with a multi-tenant database we can dynamically choose the right level that best fits our purpose.

Where we are today?

Cloud providers such as Amazon, Microsoft, and Google have already realized that they can’t expect their users to deal with multi-tenancy themselves. So they've come up with data services that were designed with multi-tenancy in mind. Amazon recently introduced Amazon RDS (Relational Database as a Service) and Microsoft introduced SQL Azure. Both are an attempt to take existing databases (MySQL in the case of Amazon and SQL Server in the case of Microsoft) and add some of the characteristics that I discussed above. For example In both cases the user works with a single tenant and is completely abstracted from the underlying resource sharing optimizations. 

And yet, both solutions are still fairy limited when it comes to providing elasticity: With the Amazon solution, you're expected to scale-up (use a bigger machine) and go through a period of downtime when you switch between one degree of scaling and another. Google AppEngine has taken a different approach. Instead of relying on an existing RDBMS and wrapping it up with multi-tenancy and dynamic scaling, they took their own NoSQL alternative (BigTable) that was already built for dynamic scaling and, through a mapping layer, provided SQL semantics on top of it. As with RDS and SQL Azure, the user is completely abstracted from the way Google manages multi-tenancy in their underlying infrastructure.

Niether Amazon, Microsoft, nor Google provide WRT flexibility to the isolation level, and more importantly, all these solutions seem to be completely tied into their own cloud data center, which means that I cannot port it to my own local environment. From a throughput efficiency perspective, SQL Azure and Google seem to trade scaling for throughput by introducing an additional layer of indirection (load balancer in the case of Azure) that makes their multi-tenant solution fit only low-end applications.

At GigaSpaces we took an approach very similar to the Google approach with our new Elastic Middleware Services, where we use our existing distributed NoSQL In-Memory Data Grid and used it as a muti-tenant data service.

That is it for similarity. Here are the differences:

• You can easily download and install GigaSpaces in your local environment or use it in the cloud and get the full multi-tenancy and dynamic scaling experience. We provide an open SPI that enables you to plug in any resource pool, such as starting from a simple SSH call through a complete virtualization pool.
• You can control the level of isolation, scaling, and other SLA directly through an API.
• Our data services are designed for extreme efficiency by maximizing the utilization of memory resources. This normally translates into the ability to run at least 10x more user transactions on the same hardware, which in turn translates into huge cost savings.
• We provide the same consistent approach for other layers of the stack such as messaging and map/reduce through the same environment.

You can learn more about the Elastic Middleware solution in my previous post.

References

• SaaS Enablement of existing applications (Videocast)

• The Missing Piece in the Virtualization Stack (Part 2)

• Why Multitenancy Matters In The Cloud by Alok Misra

• Multi Tenancy - Keep It Simple, REALLY Simple by Ayende Rahien
• Multitenancy definition – Wikipedia
• Elastic Data On the Cloud: Hype or Reality 
• GigaSpaces 7.1 Release
• Multitenant Magic - Under the Covers of the Force.com Data videocast by Craig Weisseman Chief Software Architect salesforce.com
• Security and multitenancy with GigaSpaces XAP (docs)
• Security and multitenancy with GigaSpaces XAP (videocast)

The Missing Piece in the Virtualization Stack (Part 2)

In the first part of this post, I discussed how virtualization and cloud computing, as we know it today, is only a small part of the solution for today’s IT inefficiencies. While new technologies and delivery models have made it much simpler to manage the infrastructure, this is not where our core inefficiencies lie. Virtualization principles must be extended to higher levels of the application stack, to make it easier for all of us to manage, tune and integrate applications. Otherwise we will continue to spend most of our time on things that don’t provide real value to the business – infrastructure, installation, management, tuning, and integration.

In this post, as promised, I’ll show how this missing piece can be filled in practice, by something called Elastic Middleware Services. These are middleware services that can be deployed with one API call, and are completely abstracted from the applications that use them. Elastic Middleware Services are very similar to cloud-based middleware like Amazon SimpleDB, but the major difference is that they are enterprise-grade, and can be purchased and installed in your private data center, for use by your internal enterprise applications.

To better illustrate this idea, I’ll discuss a reference implementation of Elastic Middleware Services which will shortly be released as part of GigaSpaces XAP 7.1. The discussion is intended to illustrate in concrete terms how to fill the missing piece in the virtualization stack. I’ve intentionally focused on the conceptual and fundamental elements of the GigaSpaces implementation, not on specific features and benefits, to make the discussion useful even if you are not using or considering GigaSpaces products.

By the way, I recently hosted a webinar on this topic, together with Uri Cohen, GigaSpaces’ product manager. In the webinar we presented GigaSpaces’ upcoming Elastic Middleware Services. The recorded version of this presentation is available below:

Elastic Middleware Services

The idea is to provide middleware services (messaging, data, Map/Reduce, and more), like those the ones provided by cloud providers such as Amazon and Microsoft Azure, in your own local environment. These services simply extend the basic principles of virtualization, as I outlined in the previous post, to the middleware stack:

Basic virtualization principles

1. Break big physical resources into smaller logical units
2. Decouple the application from the physical resources
3. Provide an abstraction that makes all the small units look like one big unit

From a middleware perspective, this means that instead of having one big database, you have lots of small units of data services (also called partitions or shards), which are grouped together through a client side proxy that exposes all of these small units as one big unit to the client that is using them. The client proxy is also responsible for abstracting the client from the physical location of those resources. To scale, all you need to do is add more of those units. The client experiences these additional units as extra capacity added to the service.

Deploying middleware with one API call

From an API perspective, this is a much simpler approach than traditional middleware, because it introduces a higher-level abstraction to the user, as can be seen in the code snippet below:

With this new elastic middleware API, users define requirements in a language that is closer to their domain. For example, “I want to create a data service that will have 10GB of data and can grow up to 100GB of data, I need hot failover and I'm willing to share my deployment with other users of my own organization but not with users of other organizations.”

In response, the Elastic Middleware Services would automatically do the following:

• Allocate the right number of partitions, based on machine availability and the available memory in each machine, to address the capacity required.
• Automatically launch a hot backup in for each partition and ensure that it runs on separate machine than the primary partition, to address the high-availability requirement. 
• Make sure that if another organization’s application is already running on the machine, the application won’t deploy on that machine, to address the security requirement.

Fully automated through built-in SLA

As I noted in my previous post, one of the major costs of enterprise applications over a period of three years is the cost of maintenance and operations. The main component of maintenance costs is the labor cost, closely tied to the amount of manual work you have to put into your system to make sure that it meets the application’s SLA. The cost grows proportionally to the demand for scale.

With Elastic Middleware, maintenance cost can be substantially reduced through built-in SLA and automation that covers the following aspects:

• Scaling SLA – scale when there is a memory shortage, high CPU utilization, Garbage Collector hiccups, etc. Users can choose to scale automatically when any of these events occurs, and customize their own thresholds (e.g. scale when CPU utilization hits 90%). The Elastic Middleware is integrated with a cloud or virtualization framework, to enable it to automatically pull or release of machines as needed.
• Failover – when a machine fails, the Elastic Middleware does one of two things:
  • Scale down – if there are available resources to meet the current workload, the system will automatically scale down and continue to service the application.
  • Rebalance – if there are not enough resources to serve the application, the Elastic Middleware calls the cloud/virtualization pool and starts a new machine. As soon as the new machine starts, it is added to the existing pool of resources, and the application is re-balanced to take advantage of the additional capacity available through the new machine. It is important to point out that if you don’t have a dynamic pool of resources available through virtualization or cloud computing, you could still start a machine manually. Once the machine is started, adding it to the application’s resources and rebalancing would happen in exactly the same way.
• Continuous rebalancing – when new machines are added to the Elastic Middleware, it immediately detect them and re-balances the assets currently running on the new machines (if necessary), ensuring optimum utilization of all services.

Carrying out this dynamic SLA was deigned to happen while the application is running, ensuring no transaction or data loss during the process. In-flight transactions continue to be served with no noticeable hick-ups. 

Concerned of losing control?

Whenever I introduce the concept of built-in SLA and automation, I get two type of responses. One is “Wow, that’s cool!”, the other is “It sounds like I’ll lose control.” The concern about losing control is very valid, as in many cases, especially in non-cloud environments the concept of spare capacity rarely exist and therefore adding a new machine would often involve some manual intervention. In addition, when something goes wrong, other parts of the system or even other parts of the organization need to be involved, so we can’t always assume that full automation is possible.

The cruise control analogy

The approach GigaSpaces has taken is similar to the way cruise control works in our cars. You can choose to give up full control under a certain threshold, but at each point in time you can take the wheel back and resume full control. Like many of the cloud providers, we provides this type of cruise control for our services through an open API that enables management and control of every aspect of our cluster. You can query the available node, CPU, application partitions and service components. GigaSpaces Elastic Middleware was built as a layer on top of that API, which means that if the built-in functionality that comes with the Elastic Middleware doesn’t fit your needs, you can go one layer down and write your own custom behavior. For example, you can specify that in your system, scaling or failover events send an alert that will trigger a manual process for adding a new machine. In other words, you have full flexibility to choose when to turn on “cruise control” and when to take ownership.

SaaS-enabled with built-in multi-tenancy

Multi-tenancy often means that can run multiple users/customer applications on a shared resource, thereby reducing the cost of ownership per user. This is considered an area of high complexity, specifically in SaaS applications, because with today’s middleware, the application itself is responsible for mapping between users and application tenants. With Elastic Middleware Services, multi-tenancy is built into our implementation and API. Thus the burden of sharing resources is moved from the application down to the middleware. Each application works with its own dedicated middleware service (data, messaging, etc.) but that middleware service can share resources with other middleware services. In other words, instead of having one big database running on a dedicated machine and split it out between users at the application level, you can have lots of small databases spread between machines, where each application can have its own dedicated database and still have that database shared with other database instances that runs on that same machine.

This not only significantly simplifies your code, it also provides better isolation between multiple users, as well as independent life cycle management for each tenant.

There are various tradeoffs between sharing and isolation. With isolation you get better security and control of your own environment, but lower utilization and higher cost. With sharing, you can reduce cost and still achieve reasonable isolation, but not at the same level as when running on a dedicated resource. The Elastic Middleware Services make it possible to define the isolation/sharing at the application level. It does this by introducing the notion of an isolation level:

Currently we support three isolation levels:

• Dedicated – guarantees a dedicated machine allocated per instance of the application.
• Shared private – multiple instances of the application or organization share the same resources, but other departments or organizations are isolated.
• Shared public – everyone shares everything.

 

Available both on your local network and on the cloud

To try the new service you don’t need to have a private cloud or run your application on a public cloud. You don’t even need a virtualization layer. All you need is to launch a single GigaSpaces agent (Java process) per machine (normally that process will be started automatically at machine boot).

Once you’ve done this, you can start interacting with your machines and an create the desired middleware service through a simple API call. The following snippet shows how you might use this API:

Admin admin = new AdminFactory().createAdmin();

ElasticServiceManager elasticServiceManager = admin.getElasticServiceManagers().waitForAtLeastOne();

// Start a new data-grid

ProcessingUnit pu = elasticServiceManager.deploy(new ElasticDataGridDeployment("mygrid") // give it a name

.isolationLevel(IsolationLevel.DEDICATED) //isolation level

.highlyAvailable(true) // set the high availability level

.elasticity("2GB", "6GB") // set the required capacity range

.jvmSize("512MB") // configure the size per VM

.addSla(new MemorySla(70))); // define Memory SLA

That’s it! – you just got a full cluster deployed and ready to use.

What does it mean for you?

To sum up, I here are the main benefits that Elastic Middleware Services bring to each type of user:

For developers

As a developer, you can get access to the service you want just by calling an API, without worrying about installation or cluster setup. You get access to the service from your existing platform. In other words, you don’t need to run your application on GigaSpaces XAP to use any of those services. You can also pick and choose – use only the services you want, e.g. just the data, data and messaging, etc.

For the (private) data center

From a data-center perspective, you could install Elastic Middleware Services only once in your data-center. The same installation and resources would be shared amongst all users in your organization. If you happen to have virtualization in place, you can do this very easily using the GigaSpaces virtual appliances. 

Once you install the system, other users in the organization don’t need to go through the installation process – they can start consuming the middleware services just by calling an API, like they would do on the Amazon cloud when they launch a SimpleDB instance.

In a way, the Elastic Middleware Services gives the data center the power to provide higher-lever services to the business, rather than just plain infrastructure services, enabling the organization to become more agile.

For public cloud providers

Public cloud providers can be viewed as an outsourced version of the local data center, so they should experience much the same benefits as data centers. In addition, with Elastic Middlware Services, public cloud providers can offer a set of middleware services that are enterprise-ready with extremely high utilization per service. Because the Elastic Middleware Services are fully memory-based, you can use as much as 10X less machines to achieve the same throughput requirements, compared to a disk-based approach.

Ricky Ho published some interesting performance characteristics of Amazon SQS and SimpleDB in one of his recent posts:

• Network latency and throughput: 20 - 100 ms for SQS access, SimpleDB domain write throughput is 30 - 40 items/sec.
• Eventual consistency: 2 simultaneous requests to dequeue from SQS can both get the same message. SQS sometimes reports empty when there are still messages in the queue

While both SQS and SimpleDB provides a scalable data store and messaging services, their performance are at least 10-50X slower than an equivalent GigaSpaces data and messaging service. Implementing a Map/Reduce scenario with GigaSpaces is going to be significantly simpler and closer to real-time compared to Amazon Elastic Map Reduce (and that’s a topic for a separate post). Another cool feature is the ability to execute your code within the service container, which gives you an even bigger improvement in performance and latency.

The Elastic Middleware Services have much higher utilization and simpler maintenance, compared to equivalent services offered separately to the end-user, because they are provided using a shared cluster with a unified clustering model. With Elastic Middleware Services you get JPA, memcached, key/value (NoSQL), Spring, JMS, Remoting and Map/Reduce in Java or .Net, in a single deployment.

Existing GigaSpaces XAP users

Existing GigaSpaces XAP users can use the Elastic Middleware Services as part of XAP.  They’ll benefit from the simplicity provided through the high-level abstraction. In addition, the Elastic Middleware Services make it easier to plug-in the GigaSpaces components into existing application servers or development environments. They also make it significantly simpler to manage and deploy GigaSpaces XAP in various groups in the organization, because they can all share the same virtual pool of machines in the data center.

Availability

The Elastic Middleware Services will be available as part of our upcoming XAP 7.1 release, due for end of March 2010. It is currently available for private beta testing. We will also release a public beta of the Elastic Middleware Services through our upcoming 7.1 release candidate version. As always, we’ll welcome feedback and will be very happy to hear about your specific requirements. You could either post comment to this post or send an email to <pm at gigaspaces dot com>.

Why Existing Databases (RAC) are So Breakable!

One of the core assumption behind many of today’s databases is that disks are reliable. In other words, your data is “safe” if it is stored on a disk, and indeed most database solutions rely heavily on that assumption. Is it a valid assumption?

Disks are not reliable!

Recent analysis of disk failure patterns (see references below) break the disk reliability assumption completely. Interestingly enough, even though these studies were published in 2007, most people I speak to aren’t aware of them at all. In fact, I still get comments from people who are basing their reliability solution on disk-based persistency, insisting on the argument that whether or not you are reliable amounts to whether or not your data is synched to a disk.

Below are some of the points I gathered from the research, which should surprise most of us:

• Actual disk failure/year is 3% (vs. estimates of 0.5 - 0.9%) – this is a 600% difference on reported vs. actual disk failure. The main reason is that manufacturers can’t tests their disks for years, so their tests are an extrapolation of statistical behavior which proves to be completely inaccurate.
• There is NO correlation between failure rate and disk type – whether it is SCSI, SATA, or fiber channel. Most data centers are based on the assumption that investing in high-end disks and storage devices will increase their reliability – well, it turns out that high-end disks exhibit more or less the same failure patterns as regular disks! John Mitchell had an interesting comment on this matter during our Qcon talk, when someone pointed to RAID disks as their solution for reliability. John said that since RAID is based on an exact H/W replica that lives in the same box, there is a very high likelihood that if a particular disk fails, its replica will fail in the same way. This is because they all have the exact same model, handle the exact same load and share the same power/temperature. 
• There is NO correlation between high disk temperature and failure rates – I must admit that this was a big surprise to me, but it turned out that there is no correlation between disk temperature and disk failure. In fact, most failures happen when disks were at low or average temperature.

Why are existing database clustering solutions so breakable?

Most existing database clustering solutions rely on a shared disk storage to maintain their cluster state, as can be seen in the diagram below.

 

Typical database clustering architecture based on shared storage

The core, implicit assumption behind Oracle RAC, IBM DB2, and many other database clustering solutions, is that failure can be avoided by purchasing high-end disk storage and using expensive hardware (fiber optics, etc). As can be seen from the research I mentioned earlier, this core assumption doesn’t correlate with the failure statistics. Hence I argue that the database clustering model is inherently breakable.

Failure is inevitable – cope with it

Jason McHugh, a senior engineer at Amazon who works on S3, gave a very interesting talk at the Qcon conference (Amazon S3: Architecting For Resieliency In The Face Of Failures), where he outlined the service failure handling architecture of Amazon’s Simple Storage. Jason’s main point was that rather than trying to prevent failure from happening, you should assume that failure is inevitable and instead, design your system to cope with it.

Failure can happen in various shapes and forms:

• Expect drives to fail
• Expect network connections to fail (independent of the redundancy in networking)
• Expect a single machine to go out

Human error example: Failure to use caution while pushing a rack of servers

How to design your application for resilience in the face of failure

Jason outlined some of the lessons from his experience at Amazon:

• Expect and tolerate failures – build your infrastructure to cope with catastrophic failure scenarios: "Distributed systems will experience failures, and we should design systems which tolerate rather than avoid failures. It ends up being less important to engineer components for low failure rates (MTBF) than it is to engineer components with rapid recovery rates (MTTR)." (from Architecture Summit 2008 - Write up)
• Code for large-scale failures – design your application to cope with the fact that services can come and go.
• Expect and handle data and message corruption – don’t rely only on TCP to ensure the consistency of your data. Add validation as part of your application code to make sure that messages are not corrupted.
• Code for elasticity – load/demand may grow unexpectedly. Elasticity is essential to dealing with unexpected demand without bringing the system to its knees. It is also important that when you exceed your capacity, the system won’t crash but instead report denial of service.
• Monitor, extrapolate and react – monitoring at the right level enables you to detect failures before they happen and take the appropriate corrective actions.
• Code for frequent single machine failures – rack failure can cause an immediate disappearance of multiple machines at the same time.
• Game days – simulate failures and learn how to continuously improve your operational procedure for handling failures and architecture.

Memory can be more reliable then disk

Many people assumes that memory is an unreliable data storage.  That assumption holds true if your data “lives” on a single machine; in this case if the machine fails or crashes your application crashes. But what if you distribute the data across a cluster of nodes and maintain more than one copy of the data over the network? In this case, if a node crashes the data is not gone; it lives elsewhere and can be continuously served from one of its replicas.

Under these conditions, I can argue that under certain conditions, memory can be more reliable than disk. One of the reason is the hardware itself – unlike disk drives that rely on mechanical moving parts, RAM is just a silicon chip. As such the chances for failure can be significantly smaller. Unfortunately, I don’t have the data to back up this argument, but I think it’s very likely that this is the case. In an article published on InfoQ titled RAM is the new disk, Tim Bray brings an interesting observation on this matter:

Memory is the new disk! With disk speeds growing very slowly and memory chip capacities growing exponentially, in-memory software architectures offer the prospect of orders-of-magnitude improvements in the performance of all kinds of data-intensive applications. Small (1U, 2U) rack-mounted servers with a terabyte or more or memory will be available soon, and will change how we think about the balance between memory and disk in server architectures. Disk will become the new tape, and will be used in the same way, as a sequential storage medium (streaming from disk is reasonably fast) rather than as a random-access medium (very slow). Tons of opportunities there to develop new products that can offer 10x-100x performance improvements over the existing ones.

As a matter of fact we (GigaSpaces/Cisco) are working these very days on benchmarking  Cisco UCS, where we are evaluating the possibility to manage x100GB and potentially even Terabytes of data in a single node. The results look extremely promising, so the future is not that far ahead.

References

• KHB: Real-world disk failure rates: surprises, surprises, and more surprises by Valerie Aurora (Henson)
• Disk failures in the real world: What does an MTTF of 1,000,000 hours mean to you? By Bianca Schroeder and Garth A. Gibson
• The Case for RAMClouds: Scalable High-Performance Storage Entirely in DRAM By Department of Computer Science Stanford University
  
• Amazon S3: Architecting For Resieliency In The Face Of Failures By Jason McHugh
• Failure Trends in a Large Disk Drive Population - Google, Inc., February 2007
• InfoQ Article - RAM is the new disk...
• Are Cloud Based Memory Architectures the Next Big Thing?, highscalability.com

The Interactive Cloud (Part II - Adding Operational Awareness to your Application)

In part I of of this topic, I covered the general concept and motivation behind the "Interactive Cloud". I described how the interactive cloud enables applications to be managed in a similar way to the way we run our business. In this post, I discuss further details of what it means to add operational awareness to the application. To do that, I refer to our specific experience in GigaSpaces. Towards the end of the post, I point out specific code snippets as an illustration of that model.

Background

GigaSpaces 7.0 Application Cluster Management API is an API for managing application clusters, and for giving applications operational awareness. The reason why we came up with this API in the first place, is quite interesting. When we started GigaSpaces, a typical deployment was primary-backup with multiple clients connected to this pair. Sooner then we anticipated, we found ourselves supporting cluster deployments of 10s and 100s of nodes per application. We are now at a point where customers are starting to use GigaSpaces XAP as a service shared between multiple applications, which in itself increases the deployment size quite significantly.

In this type of environment, troubleshooting a production system can be a real nightmare for everyone involved. We started to add all sorts of metrics and logging information into our system but it hardly improved things. The process of finding the right information at the level we needed was still a heavy-duty task.

When we looked back and tried to analyze how we learn about the application’s behavior, we saw that quite often we identified the time when things started to go wrong, and then, using the timestamps on the different operations, ran a complex correlation process to see what happened on each machine during that time. In many cases when we hit the first area of trouble, we were still missing information, because we were running without fine-grained logging, for performance reasons. So then we would turn up the log level and wait for the next event, and so forth.

We realized that just adding more probes and log levels and using some sort of CEP (Complex Event Processing) to parse all the garbage we were going to generate, on a continuous basis, was not going to be good enough. Even worse, if the reporting system got so complex, we would eventually end up going back to the old logs because the reporting system would inevitably lose some of our events :)

We figured that we could not afford to continue down that path – the cost, both on our end and on the customer’s end, was way too high. We needed to try a totally new approach.

The Ideal Solution – Getting Help from the Applications themselves

When we analyzed what would be the ideal process, we came to the realization that the application should be responsible for the monitoring. When the application starts to behave abnormally (for example, CPU spikes, GC spikes), or experiences network failure, it should do two things:

1. Take corrective action – the application should figure out if the problem is that resources have been exhausted. If that is the case, it should try to rebalance itself to reduce the load automatically.
2. Provide information to help analyze the root cause of the event that occurred – the application should point out the machines that are suspects and ask those server instances to increase their logging level and print a snapshot of thread dump. This provides only the information that is actually needed to perform root cause analysis, and at the right level of granularity.

It wasn’t surprising that some of the root cause analysis indicated that one of the biggest contributors to failures, is human error or misconfiguration. That led us to the realization that the way to reduce those factors would be through complete automation of manual deployment and maintenance processes.

Another thing we realized during that process, is that when things went wrong, quite often the operation guys became fairly useless at a fairly early stage of the process, and the application guys had to be called in to understand what happened. Not surprisingly, the application guys didn’t even know how to get the required information or didn’t have permission to get that information, and had to rely on someone else to do that for them – which made the entire process really long and complex.

All of this led us to the realization that we needed to create an application cluster management and monitoring tool that could be used by both the application developers and the operation guys. The application guys could use that API as part of their development cycle, and the operation guys could use it to receive the right level of information into their management and control dashboards.

In the following section, I provide a short sneak peek into this API, and how it is structured.

XAP’s New Application Cluster Management and Monitoring API (a.k.a. the Administration API)

There are basically two ways to design  the API for this challenge: define the interaction model and drive the API from there, or define the domain model and drive the interaction model from the domain model. We chose the latter.

1. Defining the Application Management Domain Model

The first step was to define the domain model of our application management. The diagram below maps out the “nervous system” of an application running on XAP – it shows the core layers of the application management domain model.

 

2. Defining the Interaction Model – the API

Once we mapped the nerves and sensors of our application, we could start to wire them. This enabled us not only to monitor the state of the application components, but also to interact with them when we needed to take corrective actions (see snippet below).

3. Making Operational Awareness an Integral Part of the Development Framework

If we want the application to take more responsibility for its operational behavior, we can’t just point it to an external management and monitoring tool that doesn’t speak the “developer language”.  We need to make the proactive and reporting behavior an integral part of the development framework. This is illustrated in the examples below.

Application Code that uses the Cluster Management API - a Few Examples

Below are a few examples that show how applications interact with the new Administration API, using a groovy-based API. These snippets were written in Groovy but they can be be called in Java as well.

Example 1 – Getting Statistics

This example uses Groovy closures to get callbacks when a new container or machine joins the network, and when a specific application event happens.

Example 2 – Calculating Average Load on a Web Container

This code snippet shows how to get an aggregated result of the total number of requests/sec from all the web containers in the cluster.

Example 3 – Scaling Up:

The code snippet below shows the process of scaling out, when the number of requests exceeds the maximum number of requests allowed per instance.

Note that the call to incrementInstance() generates an entire match-making process behind the scene. What happens is that a new web container is spawned, but the provisioning of this new container is based on the SLA that was defined at deployment time. In our case, it is deployed only on machines that don’t already run a web container and have spare CPU and memory capacity.

Initial Success Stories

Since we came up with the API, we started to see its use spreading in areas we didn’t even think of. Here are a few examples:

• Helping customers reproduce problem scenarios – now that we can script almost every part of the application, we can easily script even very complex scenarios, where the application starts to execute a few things and then crashes only when a certain operation happens, etc. The fact that we can script these scenarios makes it easier to create a reproduction scenario, and from there, the road to spotting the problem and fixing it becomes significantly shorter (see detailed reference here).
• Third part integration – one of our partners created a fairly extensible Eclipse plug-in. The reason he was able to do that pretty much by himself, was because the API gave him everything he needed in order to control the application development and testing lifecycle (see more details in Jeroen Remmerswaal post here).
• Writing a complex SLA – we were able to easily write complex SLAs, for example, automating deployment of applications across multiple data centers and making sure that primaries and backups never run on the same data center. All that without any human intervention.

What’s Coming Next?

One of the immediate things we are going to enhance, is to provide more built-in SLAs, and a set of utilities that perform deployment and implementation of mainstream scenarios, such as auto-scaling of web applications and auto re-balancing of data grid deployments. This is going to serve as a cornerstone of our new cloud-enabled middleware initiative, where all the classic, enterprise-grade middleware components become available as a service. In fact, some of our customers are already using our In-Memory Data Grid as a shared service between multiple applications within the organization. The idea is to make that type of deployment simpler and intuitive.

Take Part in this New Initiative

I can go on with more of my thoughts on what should be done, but I’d rather stop here and ask all of you for your feedback. If you’ve faced similar challenges, how do you see this solution? What would you want to see in this new package?

For those who want to try out this new API, it is worth noting that it is actually provided for free as part of our Community Edition, so anyone can simply download it and try out the examples from this post.