IBM Cell servers vs Intel Xeon Servers
I've looked for a comparison between servers running the new IBM and the Intel powerhouse processors, but the only thing I've found is the US DOE chose Cell & Opteron processors for some new supercomputer in Los Alamos.
I know you can get all three types of servers, and was wondering if anyone has had experience setting up windows server on these platforms, and if any performance difference was detected for database operations or video processing.
-Steven
Good Answers (2)
Jay B
Chasm Jumper (link to: dietcoke4u at gmail.com)
Best Answers in: Small Business (1), Starting Up (1)
First, unless you have a ton of time and budget to develop an application to take advatage of the coprocessors on the cell processor I would stick with the vanilla Intel of AMD processors.
Second, these systems have been optimized for bandwidth which is typically not a problem unless you have a giant grid. Typically latency is the killer pain the the application's behind.
Third, due to the cache coherency and bandwidth prioritization over latency these systems are KNOWN to be very difficult to program in. Ironically any performance gains due to architecture will probably be more than offset by additional cost in software development (and yes you will need lots and lots of IBM PS).
So not knowing your application but guessing that you are in HPC or grid space you might consider looking at pNFS if the issue is data bandwidth to or from the grid.
Now I will rant a bit. The fundamental problem with grid computing is NOT wiring up a bunch of hardware it is actually finding a person that understands the technical issue AND can write an application so that chunks of the code can run in parallel. Obviously the big hardware vendos want to sell lots of servers, storage, networking equipment and specialized memory to memory connectivity - yet they all ignore the fundamental problem. Maybe they should actually invest in making it easier to write grid code.
Another critical issue is that the grid code actually needs to be divided into layers for pocess management and data management. Too often the bottleneck becomes the one central database.
One last thought if you are not running a grid and just looking to distribute data looking into akamai like ECNs would be worth your time.
Rich G
Senior Product Architect at Dynamics Research Corp
Best Answers in: Computers and Software (3), Freelancing and Contracting (1), Web Development (1), Using LinkedIn (1)
Your question also asks the very specific question: "if anyone has had experience setting up windows server on these platforms"
At this time, Windows Server does not have a version for the Cell processor. While you could potentially get a multiprocessor VM configured on a Cell array, Windows Server will be running under a virtualized kernel, which will have limited CPU functionality compared to a actual hardware implementation of the same CPU.
Since you will actually be using a software copy of a CPU, you won't see any difference that would be gained, for example, by built in instructions of the Cell processor that are not implemented in the VM kernel.
Unless Microsoft comes out with a compiled copy of Windows Server, your specific question has the answer of
It would be too much work and you would see no gain whatsoever other than what clockspeed gains may be available (3.2 Ghz seems to be what the web agrees the cell runs at from a quick search).
I would suspect that at some future point, virtual machine software (such as VMWare ESX) will take advantage of the multi-pipelining in the cell processor, which will likely have noticeable gains in floating point processing, but I do not know how well those gains will transfer through the VMWare Hardware Abstraction Layer (HAL) used to emulate the Intel chipset commands for running Windows Server.
If you really want to take advantage of the Cell, right now your only option is to compile your own solutions and ensure your compiled product is optimized for multithreading your math operations.
Links:
Clarification added October 18, 2007:
By the way - I just saw a video that hints that Windows Server 2008 WILL be compiled for the Cell processor (although the vendor isn't mentioned - it's indubitably IBM) - http://www.winxpcentral.com/reviews/images/procdemo/Winprocdemo2.swf
More Answers (2)
Shannon H
Professional Projects + Product Manager at amBX ltd.
Best Answers in: Computers and Software (2), Market Research and Definition (1), Software Development (1)
The previous answer is very comprehensive. The Cell BE chips while fascinating and powerful are very challenging to work on. If you could clarify your application for super-computing a more through answer to what you should use can be generated. I highly recommend joining the top500 group here on linkedIN. And researching top500.org Also if you haven't looked into IBM's Deep Computing Sollutions it is possible to Lease dedicated Blades at IBM Deep Computing Centers as opposed to buying hardware for data centers which are expensive to run (power and cooling). Better to put money into Bandwidth. Please contact me if you would like more information and contacts to Deep Computing Specialists at IBM
Links:
Peter H
Managing Director - VeryPC - Green IT Champion
Best Answers in: Advertising (1), Green Business (1), Energy and Development (1), Computers and Software (1), Using LinkedIn (1)
The intel quad core xeon chips are so inexpensive, and also energy efficient. We've designed and manufactured a 1u unit, containing two servers each running 8 intel cores on only 112W.
With the standardisation and compatibility offered for the intel platform, it's a no brainer for HPC/Raytracing environments, especially if you add an infinband backbone.
So we're looking at 16 cores 1u (intel) vs 16 cores x 10 blades in 7u plus a headache recoding.....
I'd say the intel is a winner....
