IBM Touts Supercomputers for Enterprise 94
Stony Stevenson writes "IBM has announced an initiative to offer smaller versions of its high-performance computers to enterprise customers. The first new machine is a QS22 BladeCenter server powered by a Cell processor. Developed to power gaming systems, the Cell chip has also garnered interest from the supercomputing community owing to its ability to handle large amounts of floating point calculations. IBM hopes that the chips, which currently power climate modeling and other traditional supercomputing tasks, will also appeal to customers ranging from financial analysis firms to animation studios."
Re:Oblig.. (Score:2, Informative)
Wrong question. The correct question is: Will it run Java?
No? Then What The !@#$ have you been selling me WebFear^W^W^W^WSphere for?
BTW, for all you Eclipse lovers? NEVER install Rational Application Developer (RAD). IBM managed to take a half-decent product, add tons of suck to it, and make sure it was the SLOWEST application in the history of mankind. You need at least 4GB to run it at a decent speed. Which is pretty sad when you consider that Windows only goes up to 3.5...
Re:Flamage (Score:2, Informative)
Grrr lameness filter says I'm yelling.
Re:Oblig.. (Score:3, Informative)
Re:Oblig.. (Score:2, Informative)
Java? Who cares? (Score:5, Informative)
This is a Failure (Score:4, Informative)
I'll begin first by listing the positives:
It has very good single precision floating point performance
Very high bandwidth on chip (25 GB/s)
With the e[nhanced]D[ouble]P[recision] addition to it IBM is adding for Los Alamos' Road Runner, it should be even better.
IBM Developerworks is a useful resource for programmers
It's a processor that gives the academic community a chance to publish more papers
The negatives:
Far too expensive
Floating point performance, single or otherwise, is useless for most enterprise work (no point running a database on a Cell)
The things that makes the Cell fast and unique are the SPEs (SIMD processors) are useless for servers*(I'll elaborate below)
Developing software to make use of the SPEs is time consuming and difficult. Orders of magnitude more so if your algorithm isn't suited for being split across SPEs
Computer Science or Engineering students who are doing research in this field have to read redundant papers that reword the IBM Cell manual (plagarism?) for 70% of it and what they actually did with the Cell for about 10%. 20% is of course left for citations.
You can't actually play games on a Cell. The PS3 games use them for physics engines, sound or such; the video is done on a video card.
Worst of all, YOU HAVE TO GET DATA FROM MAIN MEMORY TO SPU'S CACHE YOURSELF! (also known as Local Store, 256KB)
Now to talk about the SPEs... They are what makes the Cell tick. If the Cell didn't have them, it would just be an old Mac processor. (IBM Power 4 was it? I've trying to avoid research papers on the Cell) The graduate students (a year away from a PhD) that I was observing who were doing developing a scientific application on the Cell.. when we summed their year in development, they essentially told me it was a pain in the ass. Why? Because you have to get data from main memory to local store. Imagine if you had to get data from memory to your x86 processors L2 or L1 cache. (No, you don't actually want that. Trust me, you don't.) Scientists don't want to develop applications on this platform. Researchers like it because they can publish papers talking about it. This is why Los Alamos' Road Runner might be a flop. Bye bye to $100m of US tax payer money.
Back to the discussion with the grad students... I then came up with the idea that the Cell would be perfectly suited to an Asterix server. Why? Because sometimes Asterix needs to convert from one format to another when audio codecs differ. This is a perfect SIMD application, except we have a problem. While the Cell does have great on chip bandwidth, it has very poor Power5 processor (PPU) to SPU latency*. There is a research paper out there that puts PPU initiated memory transfers to SPU latency at about 4 microseconds. Compare this to SPU initiated transfer to SPU latency of about a tenth of a microsecond. Huge difference. You can't really avoid this unless you want to develop something horrendous that avoid the PPU but that's unlikely. The PPU would run the actual server and then it would need to notify the SPU of data, either by sending message (slow) and the SPU getting the data itself (2*slow=2slow) or sending the data and a message (slower).
It will be about 5 years until good software development tools are written for this architecture that will address most of the issues. This is assuming that IBM doesn't ditch it, which seemed likely before this announcement.
Oh, by the way, the QS22 doesn't support a hard drive which means you need fast NFS. The cost of ownership is more than just the cost of single blade.
Re:The trend towards commodity hardware continues. (Score:3, Informative)
Don't get me wrong, it definately had its uses and we do use it some where I work, but it's not the be-all end-all that VMWare and co. would like you to believe. Basically if you're not using it for:
a) Test environments
or
b) Services + apps that don't play well with others
then it is probably a waste of resources. Like you said, running a virtual machine for just an FTP server is absurd...."Let's use 50x the ram and hard drive space necessary for an FTP server!"