Monday, April 25, 2005

Application Servers on Atkins

The trend towards lightweight servers has clearly moved beyond the LAMP open source stack into the enterprise Java space as well. More and more new Java deployments run on clusters of Apache Tomcat or on the web containers of J2EE servers. EJB 3 will introduce EntityBeans with POJOs (plain old Java objects), so the J2EE standard will finally catch up with popular open source technologies like Hibernate (although of course unnecessarily reinventing it instead of just using the popular open source that is already established). I think we can safely say that the trend in both the LAMP and Java worlds favors lightweight servers running on commodity hardware, rather than small clusters of heavyweight servers running on SMP hardware.

Unfortunately I can not attribute these quotes, but an SVP at IBM shared with me that he knows of only five WebSphere customers that are coordinating multiphase commits from WebSphere. A technologist I know at BEA knew of NO customers coordinating multiphase commits from WebLogic. For the bit-twiddlers out there, I mean using JTA/JTS. XA or a message queue both work just as well from lightweight servers.

Clearly the pain point for the majority of middle tier applications is throughput, especially as more and more web services come online. In short, workloads that accept a lot of connections from clients and make a lot of connections to backends, where systems need to deliver high levels of throughput computing around HTML and XML, rather than traditional processor intensive computing.

On the polar opposite of this trend towards lightweight servers are companies like Azul Systems that are staunchly in the heavyweight server camp. Azul has invested a rumored $100 million on a proprietary software/hardware appliance system for J2EE applications that need to do a lot of processing other than throughput. The Azul appliance is used by existing J2EE clusters to offload compute-intensive processing. The Azul appliance does not have any application level I/O capabilities, and uses the host J2EE cluster to connect to databases and backend systems, so even things like writing to a log file needs to go back to the J2EE cluster. In my opinion there aren't too many applications like this beyond heavy duty ERP and supply chain systems, and those are already over-provisioned with big honking Sun and IBM boxes.

But here's the really funny part: the Azul hardware runs on their custom "Vega" cores, 24 of which fit on a chip, which I am guessing are roughly equivalent in processing power to a MIPS or ARM core. We designed similar systems when I was at SunLabs, so I am speaking from experience here. There is a reason we didn't end up building them, the math didn't add up in terms of price/performance compared to the AMD and Intel roadmaps.

Here is a back-of-the-envelope calculation for Azul's price/performance. Azul's 16 chip, 384 core machine with 256GB of RAM costs $800,000, or $2,083 per core with 680MB RAM. Each Azul core looks like it is barely more powerful than a low power RISC chip from a high-end PDA. At the system level, each $2,083 Azul core costs more than a dual processor Xeon system, which can smoke a low power RISC core.

It would be great to do a real price/performance comparison, but this "fundamentally new approach to delivering compute resources" for "unbound compute" unfortunately is the first new chip in my recollection that shipped with no SPECint numbers. Nor are there any SPECjbb or ECperf numbers. J2EE is a very mature platform with well known benchmarks that can be used to show how much better a particular platform is. Clearly there is a reason for the lack of metrics here. :)

The bottom line is that there is no software/hardware stack on the planet more vertically optimized than the Apache/Linux/x86/GigE "commodity stack", and everything from LAMP to Java to animation rendering will run on it. Anything that tries to compete with the commodity stack is going to have a tough, tough time. Anything that builds on top of the commodity stack is riding the wave to the future.

Monday, April 11, 2005

The Bits are Over the Wall!

I am excited to announce that we have shipped the early access release of the ActiveGrid Grid Application Server and ActiveGrid Application Builder, which are shipping under the Apache 2 license. It took almost two years to get here, and we feel that the early access is a solid foundation on which we can evolve ActiveGrid's technology. The early access release is focused on the development and deployment of basic database-oriented web applications. We will be shipping regular releases through our 1.0 release which is targeted for the June/July timeframe. The 1.0 release will include CSS style editors, DHTML and JavaScript support, a full BPEL editor, and full web services integration.

We have made a concerted effort to ensure that (at least on Windows) downloading the ActiveGrid Application Builder and getting started is a simple and painless experience. It should take no more than 15 minutes to install the Application Builder, run the included demos and start building applications that tie into your existing backends.

To download the release, go to:

Download Early Access

To check out screenshots, go to:

Screen Shots

For our DeveloperGrid developer community, go to:

DeveloperGrid

We hope that you enjoy checking out ActiveGrid and consider joining our community as either a contributing user or as a developer. Let us know what you think, the good, the bad, and the ugly!