According to eWEEK and HPCWire, a U.K. startup company, Omneta plc, is about to build a distributed supercluster able to reach a peak performance of 120 teraflops for a sustained performance of about 70 teraflops. This supercomputer, which would become the most powerful in Europe, will be made of several thousands of Apple Xserve G5 servers and Xserve RAID storage systems. The cluster will use the Interoute network which covers today 61 cities in 19 countries. The system should be commercially available next year, offering several solutions to potential customers, such as pay-as-you-go programs or fixed term contracts. However, there are significant issues to solve before this deployment can be successful, both financial and technical.
First, here is a small diagram of the current European Interoute network map, covering more than 23,000 kilometers in 19 countries (Credit: Interoute). You can download larger versions of this map in several formats from this page if you want to have a better idea of the cities linked by this network.
What are the difficulties facing such a project?
- Let's start with the funding. Omneta has been funded with about $4 million, which is not much considering the cost of the processors alone. An Apple Xserve G5 today has two PowerPC G5 processors running at 2.3 GHz and delivering each 4 FLOPS per clock cycle. This gives the Xserve G5 a peak performance of 18.4 gigaflops. In order to reach 120 teraflops, the system will need to be made from more than 6,000 servers which are priced individually around $3,000. Even with a large discount from Apple, the servers alone should cost more than the money that Omneta has in the bank. And obviously, there are other hardware costs such as the storage systems and the network.
- After the system is built, the company will need to find customers. Omneta CEO Steven Windmill said to HPCWire: "Omneta will make supercomputer power easily available -- enabling doctors to rapidly analyze data that can save lives. Banks will also be able to monitor and detect money laundering and insider trading on a real-time basis; potentially reducing the threat of terrorism." But will these potential customers will be convinced and will their applications run on this cluster?
- Then, there are the processors. Apple can certainly supply more than 6,000 servers, equipped with PowerPC G5 processors, but until when? I've not recently seen a roadmap for the Xserve G5, but with the future move of Apple to Intel chips, what will be the processors inside these servers in a couple of years? This has the potential to seriously harm the hardware scalability of this cluster.
- And of course, the Interoute network will need to deliver a pretty impressive bandwidth. Omneta envisions a capacity to manage over a petabit per second (a billion megabits per second) of traffic at multi-teraflop peak performance. Can the Interoute network such a level of performance?
- There are also other scalability issues associated with the operating system. Speaking to eWEEK, Omneta chief technical office Tony Clark said, "According to Apple, theoretically, the Tiger OS is infinitely scalable. 'Theoretically,' in plain English, means that they don't know." So even Omneta CTO doesn't seem so sure that the Apple OS X can scale to more than 12,000 processors.
- Finally, there are security issues as well: authentication is one, but delivering data to the wrong customer is another serious issue with such an architecture. Here is another quote from the eWEEK article. Omneta will also be able to assure customers that the network will not share information outside of certain geographic boundaries. For example, a Swiss pharmaceutical company which might be concerned that its data might be stored or transmitted into a server elsewhere, Clark said."
I have no doubt that the engineers from Omneta, Apple, Interoute and their partners have enough technical expertise and enthusiasm to successfully build this supercluster, but so many things can go wrong in such a project that I'm wishing them good luck.
So what do you think of this future supercluster? Can it be built and be a commercial success? And how do you see its future?
Sources: Mark Hachman, eWEEK, September 21, 2005; HPCWire, September 23, 2005; and various web sites
You'll find related stories by following the links below.