The Pattern Bible for Distributed Computing

Software design patterns are an emerging tool for guiding and documenting system design. Patterns usually describe software abstractions used by advanced designers and programmers in their software. Patterns can provide guidance for designing highly scalable distributed systems. Let's see how! Patterns are in essence solutions to problems. Most of them are expressed in a format called Alexandrian form which draws on constructs used by Christopher Alexander. There are variants but most look like this:

  • The pattern name
  • The problem the pattern is trying to solve
  • Context
  • Solution
  • Examples
  • Design rationale: This tells where the pattern came from, why it works, and why experts use it
Patterns rarely stand alone. Each pattern works on a context, and transforms the system in that context to produce a new system in a new context. New problems arise in the new system and context, and the next ‘‘layer’’ of patterns can be applied. A pattern language is a structured collection of such patterns that build on each other to transform needs and constraints into an architecture. The latest POSA book Pattern-Oriented Software Architecture Volume 4: A Pattern Language for Distributed Computing will guide the readers through the best practices and introduce them to key areas of building distributed software systems using patterns. The book pulls together 114 patterns and shows how to use them in the context of distributed software architectures. Although somewhat theoretical it is still a great resource for practicing distributed-systems architects. It is as close as you're going to get to a one-stop "encyclopedia" of patterns relevant to distributed computing. However it is not a true encyclopedia since "over 150" patterns are referenced but not described in POSA Volume 4. The book does not go into the details of the pattern's implementations, so the reader should already be familiar with the patterns, or be prepared to spend some time researching. The pattern language for distributed computing includes patterns such as:
  • Broker
  • Client-Dispatcher-Server
  • Pipes and Filters
  • Leaders/Followers
  • Reactor
  • Proactor
Patterns can indeed be useful in designing highly scalable systems and solving various problems related to concurrency, synchronization and resource management and other topics. Wikipedia has more details on Pattern languages to check out.

Click to read more ...


Scalability Worst Practices

Brian Zimmer, architect at travel startup Yapta, highlights some worst practices jeopardizing the growth and scalability of a system: * The Golden Hammer. Forcing a particular technology to work in ways it was not intended is sometimes counter-productive. Using a database to store key-value pairs is one example. Another example is using threads to program for concurrency. * Resource Abuse. Manage the availability of shared resources because when they fail, by definition, their failure is experienced pervasively rather than in isolation. For example, connection management to the database through a thread pool. * Big Ball of Mud. Failure to manage dependencies inhibits agility and scalability. * Everything or Something. In both code and application dependency management, the worst practice is not understanding the relationships and formulating a model to facilitate their management. Failure to enforce diligent control is a contributing scalability inhibiter. * Forgetting to check the time. To properly scale a system it is imperative to manage the time alloted for requests to be handled. * Hero Pattern. One popular solution to the operation issue is a Hero who can and often will manage the bulk of the operational needs. For a large system of many components this approach does not scale, yet it is one of the most frequently-deployed solutions. * Not automating. A system too dependent on human intervention, frequently the result of having a Hero, is dangerously exposed to issues of reproducibility and hit-by-a-bus syndrome. * Monitoring. Monitoring, like testing, is often one of the first items sacrificed when time is tight.

Click to read more ...


Product: Happy = Hadoop + Python

Has a Java only Hadoop been getting you down? Now you can be Happy. Happy is a framework for writing map-reduce programs for Hadoop using Jython. It files off the sharp edges on Hadoop and makes writing map-reduce programs a breeze. There's really no history yet on Happy, but I'm delighted at the idea of being able to map-reduce in other languages. The more ways the better. From the website:

Happy is a framework that allows Hadoop jobs to be written and run in Python 2.2 using Jython. It is an 
easy way to write map-reduce programs for Hadoop, and includes some new useful features as well. 
The current release supports Hadoop 0.17.2.

Map-reduce jobs in Happy are defined by sub-classing happy.HappyJob and implementing a 
map(records, task) and reduce(key, values, task) function. Then you create an instance of the 
class, set the job parameters (such as inputs and outputs) and call run().

When you call run(), Happy serializes your job instance and copies it and all accompanying 
libraries out to the Hadoop cluster. Then for each task in the Hadoop job, your job instance is 
de-serialized and map or reduce is called.

The task results are written out using a collector, but aggregate statistics and other roll-up 
information can be stored in the happy.results dictionary, which is returned from the run() call.

Jython modules and Java jar files that are being called by your code can be specified using 
the environment variable HAPPY_PATH. These are added to the Python path at startup, and 
are also automatically included when jobs are sent to Hadoop. The path is stored in happy.path 
and can be edited at runtime. 

Click to read more ...


Lucasfilm: The Real Magic is in the Data Center

Kevin Clark, director of IT operations for Lucasfilm, discusses how their data center works: * Linux-based platform, SUSE (looking to change), and a lot of proprietary open source applications for content creation. * 4,500-processor render farm in the datacenter. Workstations are used off hours. * Developed their own proprietary scheduler to schedule their 5,500 available processors. * Render nodes, the blade racks (from Verari), run dual-core dual Opteron chips with 32GB of memory on board, but are expanding those to quad-core. Are an AMD shop. * 400TB of storage online for production. * Every night they write out 10-20TB of new data on a render. A project will use up to a hundred-plus terabytes of storage. * Incremental backups are a challenge because the data changes up to 50 percent over a week. * NetApps used for storage. They like the global namespace in the virtual file system. * Foundry Networks architecture shop. One of the larger 10-GbE-backbone facilities on the West coast. 350-plus 10 GbE ports that used for distribution throughout the facility and the backend. * Grid computing used for over 4 years. * A 10-Gig dark fiber connection connects San Rafael and their home office. Enables them to co-render and co-storage between the two facilities. No difference in performance in terms of where they went to look for their data and their shots. * Artists get server class machines: HP 9400 workstations with dual-core dual Opteron processors and 16GB of memory. * Challenge now is to better segment storage to not continue to sink costs into high-cost disks. * VMware used to host a lot of development environments. Allows the quick turn up of testing as the tests can be allocated across VMs. * Provides PoE (power-over-ethernet) out from the switch to all of our Web farms. * Next push on the facilities side. How to be more efficient at airflow management and power utilization.

Click to read more ...

Sep252008 Rated 16th Best Blog for Development Managers

Jurgen Appelo of asked for nominations for top blogs and then performed a sophisticated weighting of their popularity based on page range, trust authority, Alexa rank, Google hits, and number of comments. When all the results poured out of the blender HighScalability was ranked 16th. Not bad! Joel on Software was number one, of course. The next few were: 2) Coding Horror by Jeff Atwood 3) Seth's Blog by Seth Godin and 4) Paul Graham: Essays Paul Graham. All excellent blogs. Very cool.

Click to read more ...


Is your cloud as scalable as you think it is?

An unstated assumption is that clouds are scalable. But are they? Stick thousands upon thousands of machines together and there are a lot of potential bottlenecks just waiting to choke off your scalability supply. And if the cloud is scalable what are the chances that your application is really linearly scalable? At 10 machines all may be well. Even at 50 machines the seas look calm. But at 100, 200, or 500 machines all hell might break loose. How do you know? You know through real life testing. These kinds of tests are brutally hard and complicated. who wants to do all the incredibly precise and difficult work of producing cloud scalability tests? GridDynamics has stepped up to the challenge and has just released their Cloud Performance Reports. The report is quite detailed so I'll just cover what I found most interesting. GridDynamics in this report test three configurations:

  • GridGain running a Monte-Carlo simulation on EC2. This test is a CPU only test, a data grid is not accessed. This scenario tests the scalability of EC2 and GridGain. * GridGain provided near linear scalability end-to-end on a 512 node EC2 hosted grid. * EC2 is ready for production usage on large-scale stateless computations exhibiting good price for performance and a strong linear scaling curve.
  • GigaSpaces running a risk management simulation on EC2. This is a data-driven test. GigaSpaces is used in a configuration where the compute grid and the data grid are separated, even though GigaSpaces supports an in-memory data grid. * GigaSpaces provided near linear scalability from 16 to 256 nodes. There was a 28% degradation from 256 to 512 nodes because only four data grid servers were used. More were needed. The compute grid and data grid must each be sized to independently to scale properly. * EC2 is ready for production usage for classes of large-scale data-driven applications.
  • Windows HPC Server and Velocity running an analytics application in Microsoft's grid testbed. This test was more complicated than the others. It tested the performance implications of data "in the cloud" vs. "outside the cloud" for data-intensive analytics applications. * Keeping data close to the business logic matters. Performance improved up to 31 times over "outside the cloud." * Velocity failed on 50 node clusters with 200 concurrent clients. * Local caches provided significant performance gains over distributed caches. The local cache took load off the distributed cache. They are currently running more tests with different configurations. Hopefully we'll see those results later. All-in-all a generally optimistic report. EC2 scales. Mot of the tested grid frameworks scaled. What's also clear is it may take a while before deploying cloud based grids is an easy process. It still takes a lot of work to install, configure, start, stop, monitor, and debug bottlenecks in cloud based grids. Thanks to GridDynamics for putting in all this work and I look forward to their next set of reports.

    Click to read more ...

  • Thursday

    GridGain: One Compute Grid, Many Data Grids

    GridGain was kind enough to present at the September 17th instance of the Silicon Valley Cloud Computing Group. I've been curious about GridGain so I was glad to see them there. In short GridGain is: an open source computational grid framework that enables Java developers to improve general performance of processing intensive applications by splitting and parallelizing the workload. GridGain can also be thought of as a set of middleware primitives for building applications. GridGain's peer group of competitors includes GigaSpaces, Terracotta, Coherence, and Hadoop. The speaker for GridGain was the President and Founder, Nikita Ivanov. He has a very pleasant down-to-earth way about him that contrasts nicely with a field given to religious discussions of complex taxomic definitions. Nikita first talked about cloud computing in general. He feels Java is the perfect gateway for cloud computing. Which is good because GridGain only works with Java. The Java centricity of GridGain may be an immediate deal killer or a non-issue for a Java shop. Being so close to the language does offer a lot of power, but it sure sucks in a multi-language environment. Nikita gave a few definitions which are key to understanding where GridGain stands in the grid matrix:

  • Compute Grids: parallel execution.
  • Data Grids: parallel data storage.
  • Grid Computing: Compute Grids + Data Grids
  • Cloud Computing: datacenter + API. The key is automation via programmability as a way to deploy applications. The advantage is a unified programming model. Build an application on one node and you can run on many nodes without code change. Moving peak loads to the cloud can give you a 10x-100x cost reduction. Cloud computing poses a number of challenges: deployment, data sharing, load balancing, failover, discovery (nodes, availability), provisioning (add, remove), management, monitoring, development process, debugging, inter and external clouds (syncing data, syncing code, failover jobs). Nakita talked some about these issues, but he didn't go in-depth. But he showed a good understanding of the issues involved so I would be inclined to think GridGain handles them well. The cloud computing section is new to the standard GridGain presentation. GridGain is moving their grid into the cloud with new features like a cloud management layer available in Q1 2009. This move competes with GigaSpaces early move to the cloud with their RightScale partnership. It's a good move. Like peanut butter and chocolate, grids and clouds go better together. Grids have been under utilized largely because of infrastructure issues. A cloud platform makes it is to affordably grow and manage grids, so we might see an uptick in grid adoption as clouds and grids hookup. GridGain positions themselves as a developer centric framework according to their analysis of cloud computing in Java:
  • Heavy UI oriented. These types of applications or framework usually provide UI-based consoles, management applications, plugins, etc that provide the only way to manage resources on the cloud such as starting and stopping the image, etc. The key characteristic of this approach is that it requires a substantial user input and human interaction and thus they tend to be less dynamic and less on-demand. Good examples would be RightScale, GigaSpaces, ElasticGrid.
  • Heavy framework oriented. This approach strongly emphasizes dynamism of resource management on the cloud. The key characteristic of this approach is that it requires no human interaction and all resource management can be done programmatically by the grid/cloud middleware - and thus it is more dynamic, automated and true on-demand. Google App Engine (for Python), GridGain would be good examples. I think there's a misunderstanding of RightScale here. The UI is to configure the automated system, not manage the system. The automated system monitors and responds to events without human interaction. Won't their automated cloud layer have to do something similar? To bootstrap any complex system out of the mud of complexity a helpful UI is needed. The framework approach of GridGain's infrastructure is developer friendly, but that won't fly for external management within the cloud.

    GridGain's True Nature: One Compute Grid, Many Data Grids

    With these definitions in place we can now learn the secret of Grid Gain: One Compute Grid, Many Data Grids. Ding! Ding! Ding! Once I understood this I understood Grid Gain's niche. GridGain has focussed on making it dead simple to distribute work across a compute grid. It's a job management mechanism. GridGain doesn't include a data grid. It will work against any data grid. For some reason this fact was something I'd never pulled out of the noise before. And when I would read Nakita's blog with all the nifty little code samples I never really appreciated what was happening. Yes, I'm just that dumb, but I also think Grid Gain should expose the magic of what's going on behind the scenes more rather than push the simple 30-second-lets-write-code-live style demo. Seeing the mechanics would make it easier to build a mental model of the value being added by GridGain.

    Transparent and Low Configuration Implementation of Key Features

    A compute grid is just a bunch of CPUs calculations/jobs/work can be run on. As a developer problem are broken up into smaller tasks and spread across all your nodes so the result is calculated faster because it is happening in parallel. GridGain enthusiastically supports the MapReduce model of computation. When deploying a grid a few key problems come up:
  • How do you get your code to all nodes? Not just the first time, but every time a JAR file changes how distributed across all nodes?
  • How do all the other nodes find each other when they come up? Clearly for work to be sent to nodes someone must know about them.
  • How are jobs distributed to the nodes? Somehow jobs must be sent to a node, the calculations made, and the results assembled.
  • How are failures handled? Somehow when a node goes down and new nodes come on-line work must be rescheduled.
  • How does each node get the data it needs to do its work? Scalable computation without scalable data doesn't work for most problems. Much of the drama is lost with GridGain because most of these capabilities almost are implemented almost transparently. Discovery happens automatically. When nodes come up they communicate with each other and transparently form a grid. You don't see this, it just happens. In fact, this was one of GridGain's issues when porting to the cloud. They used multicast for discovery and Amazon doesn't support multicast. So they had to use another messaging service, which GridGain supports doing out-of-the box, and are now working on their unicast own version of the discovery service. Deploying new code is always a frustrating problem. Over the same transparently formed grid, code updates are transparently auto deployed on the grid. Again, this is one of those things you see happen from Eclipse and it loses most of the impact. It just looks like how it's supposed to work, but rarely does. With GridGain you do a build and your code changes are automatically sent through to each node in the grid. Very nice. To mark a method a gridified an annotation (or an API call) is used:
    @Gridify(taskClass = GridifyHelloWorldTask.class, timeout = 3000)
    public static int sayIt(String phrase) {
        // Simply print out the argument.
        System.out.println(">>> Printing '" + phrase + "' on this node from grid-enabled method.");
        return phrase.length();
    The task class is responsible for splitting method execution into sub-jobs. For a full example go here. The @Gridify annotation uses AOP (aspect-oriented programming) to automatically "gridify" the method. I assume this registers the method with the job scheduling system. When the application comes up and triggers execution the method is then scheduled through the job scheduling system and allocated to nodes. Again, you don't see this and they really don't talk enough about how this part works. Notice how so much complexity is nicely hidden by GridGain with very little configuration on the developer's part. There aren't a billion different XML files where every single part of the system has to be defined ahead of time. The dynamic transparent nature of the core features make it simple to use.

    Integrating with the Data Grid

    We haven't talked about data at all. If you are just concerned with a program like a Monte Carlo simulation then the compute grid is all you need. But most calculations require data. Where does your massive compute grid pull the data from? That's where the data grid comes in. A data grid is the controlled sharing and management of large amounts of distributed data. GridGain leaves the data grid up to other software by integrating with packages like, JBoss Cache, Oracle Coherence, and GigaSpaces. Remember One Compute grid, Many Data Grids. GridGain accesses the data grid through an API so you can plug in any data grid you want to support with a little custom code. Google and Hadoop use a distributed file system (DFS) as their data grid. This makes sense. When you need to feed lots of CPUs the data can't come from a centralized store. The data must be parallelized and that's what a DFS does. A DFS splats data across a lot of spindles so it can be pulled relatively quickly by lots of CPUs in parallel. Other products like Coherence and GigaSpaces store data in an in-memory data grid instead of a filesytem. Serving data from memory is faster, but you are limited by the amount of memory you have. If you have a petabyte of data your choice is clear, but if your problem is a bit smaller than maybe an in-memory solution would work. The closer data is to the business logic the better performance will be. GridGain controls job execution while the data grid is responsible for the availability and integrity of the data. GridGain doesn't care what data grid you use, but your choice has implications for performance. A compute grid and an in-memory data grid in the same cloud will smoke configurations where the data grid comes from disk or is located outside the cloud.

    GridGain is Linearly Scalable for a Pure CPU Benchmark

    The good folks at GridDynamics are doing some serious cloud testing of different products and different clouds. They did a test Scalability Benchmark of Monte Carlo Simulation on Amazon EC2 with GridGain Software that found GridGain was linearly scalable to 512 nodes in Amazon's EC2. A Monte Carlo simulation is a CPU test only, it does not use a data grid. A data grid based test would be more useful to me as everything changes once large amounts of data start flying around, but it does indicate the core of GridGain is quite scalable.

    Wrapping Up

    Grid products like Coherence and GigaSpaces include both compute grid and data grid features. Why choose a compute grid only system like GridGain when other products include both capabilities? GridGain might say they win business on the quality of their compute grid, excellent support and documentation, and the ability to cleanly integrate into almost any existing ecosystem through their well thought out API abstraction layer and their out-of-the-box support for almost every important Java framework. Others may counter performance is far better when the business logic and the job management are integrated. All interesting issues to tradeoff in your own decision making process. GridGain is free as their business model is based on providing support and consultation. A non-starter for many is the Java-only restriction. What is unique about GridGain is how easy and transparent they made it to use and deploy. That's some thoughtful engineering.

    Related Articles

  • Gridify Blog
  • Ten Useful Gridgain How-To Tips
  • 10 Reasons to Use GridGain
  • What is Grid Gain?
  • Developers Productivity: Unsung Hero of GridGain
  • GridGain vs Hadoop
  • Cameron Purdy: Defining a Data Grid
  • Compute Grids vs. Data Grids

    Click to read more ...

  • Wednesday

    Building a Scalable Architecture for Web Apps

    By Bhavin Turakhia CEO, Directi. Covers: * Why scalability is important. Viral marketing can result in instant success. With RSS/Ajax/SOA number of requests grow exponentially with user base. Goal is to build a web 2.0 app that can server millions of users with zero downtime. * Introduction to the variables. Scalability, performance, responsiveness, availability, downtime impact, cost, maintenance effort. * Introduction to the factors. Platform selection, hardware, application design, database architecture, deployment architecture, storage architecture, abuse prevention, monitoring mechanisms, etc. * Building our own scalable architecture in incremental steps: vertical scaling, vertical partitioning, horizontal scaling, horizontal partitioning, etc. First buy bigger. Then deploy each service on a separate node. Then increase the number of nudes and load balance. Deal with session management. Remove single points of failure. Use a shared nothing cluster. Choice of master-slave multi-master setup. Replication can be synchronous or asynchronous. * Platform selection considerations. Use a global redirector for serving multiple datacenters. Add object, session API, and page cache. Add reverse proxy. Think about CDNs, Grid computing. * Tips. Use a Horizontal DB architecture from the beginning. Loosely couple all modules. Use a REST interface for easier caching. Perform application sizing ongoingly to ensure optimal hardware utilization.

    Click to read more ...


    How to Scale with Ruby on Rails

    By George Palmer of Covers: * How you start out: shared hosting, web server DB on same machine. Move two 2 machines. Minimal code changes. * Scaling the database. Add read slaves on their own machines. Then master-master setup. Still minimal code changes. * Scaling the web server. Load balance against multiple application servers. Application servers scale but the database doesn't. * User clusters. Partition and allocate users to their own dedicated cluster. Requires substantial code changes. * Caching. A large percentage of hits are read only. Use reverse proxy, memcached, and language specific cache. * Elastic architectures. Based on Amazon EC2. Start and stop instances on demand. For global applications keep a cache on each continent, assign users to clusters by location, maintain app servers on each continent, use transaction replication software if you must replicate your site globally.

    Click to read more ...


    The 7 Stages of Scaling Web Apps

    By John Engales CTO, Rackspace. Good presentation of the stages a typical successful website goes through: * Stage 1 - The Beginning: Simple architecture, low complexity. no redundancy. Firewall, load balancer, a pair of web servers, database server, and internal storage. * Stage 2 - More of the same, just bigger. * Stage 3 - The Pain Begins: publicity hits. Use reverse proxy, cache static content, load balancers, more databases, re-coding. * Stage 4 - The Pain Intensifies: caching with memcached, writes overload and replication takes too long, start database partitioning, shared storage makes sense for content, significant re-architecting for DB. * Stage 5 - This Really Hurts!: rethink entire application, partition on geography user ID, etc, create user clusters, using hashing scheme for locating which user belongs to which cluster. * Stage 6 - Getting a little less painful: scalable application and database architecture, acceptable performance, starting to add ne features again, optimizing some code, still growing but manageable. * Stage 7 - Entering the unknown: where are the remaining bottlenecks (power, space, bandwidth, CDN, firewall, load balancer, storage, people, process, database), all eggs in one basked (single datacenter, single instance of data).

    Click to read more ...