Prime95, hyperthreading, multiple processors, Win2003, etc...

[pcr]

A server with dual P4 2.8GHz hyperthreaded Xeon processors (running Win2003 Server) has come available to me, and so I just threw P95 24.14 on it. Windows sees this configuration as 4 processors; 2 physical, each with 2 logical processors. When Prime95 started up, I took a look at the Task Manager and saw that the CPU utilization, while stuck firmly at 25% (as expected with a single-threaded process on a machine with 4 processors), utilization was jumping up and down on each processor. I take this to mean that Prime95 is bouncing between all the processors.

Apparently, Windows' logic for assigning threads to processors doesn't have much inherent processor affinity built-in. Somebody jump in here and tell me if I've gone off the rails, but I would think that the operating system ought to try to keep a thread on one processor as much as possible, in order to take advantage of memory caching. By bouncing the thread between processors, wouldn't we lose performance because we'd keep dumping the program from the cache?

Anyway, I went to the P95 Advanced menu and selected a single processor on the Affinity menu. Oddly enough, that reduced the task bouncing, but it didn't eliminate it. So, I turned off Affinity in the P95 program and then used Task Manager to assign Prime95 to processor 1. That stopped the bouncing.

On my single-processor P4s with hyperthreading, I haven't bothered to set processor affinity for Prime95, since I figure on a single processor machine, both logical processors are executing out of the same cache, so there'd be no benefit.

Am I on the right track here?

[moo]

rightclick on the process and chose what procs it should run on default is all of them. also set it with prime95

[JHagerson]

It appears that the least performance degradation is to assign one LL test and one factoring assignment to each physical processor. You don't get full doubling, but it appears that the contention for the FPU is minimized.

As to your contention that a task should have affinity to one processor to maximize the cache, I would agree with you. Microsoft claims to want to run on big iron, but they don't yet seem to define "big iron" as having more than one Intel processor in it. (Please don't flame me for not mentioning AMD. In the "old days," Microsoft developed on and for Compaq computers with Intel processors and Hewlett-Packard printers. Anyone else who chose to be compatible with the HPIntelpaq hardware could climb on board.)

To me, "big iron" is capable of running a bank or a brokerage firm. Databases that are up 24 hours a day for YEARS at a time handling hundreds of agents and thousands of transactions per minute. Microsoft seems more focused on running your DVD player and making sure that you can't do "too much" with it. Then, you need to reboot your computer at least once every month to install the latest update. Didn't the IBM 370 have the software update problem solved in the 60s? Maybe it wasn't until VM/MVS, but IBM solved it and Microsoft has not.

[moo]

Thats because microsoft doesnt think its customors are worthwile.... Give them an island and they will call themselves a superpower....

[xilman]

Originally Posted by JHagerson

As to your contention that a task should have affinity to one processor to maximize the cache, I would agree with you. Microsoft claims to want to run on big iron, but they don't yet seem to define "big iron" as having more than one Intel processor in it. (Please don't flame me for not mentioning AMD.

That statement is complete drivel and has been for years. Try researching what Windows server operating systems can really do before making such pronouncements.

Microsoft have been shipping and using multiprocessor-capable operating systems for years. My first toy on joining Microsoft Research in early 1998 was a quad-processor Compaq box (four 200MHz PPro cpus to be precise). In May 98 I and a colleague built a cluster from 16 dual PII-300 machines. Four years later I built its successor, again 16 dualprocs but 1GHz PIII this time.

Don't be fooled by the prevalence of toy computers running toy operating systems. Just because that's where the most money lies doesn't mean that that's all Microsoft produces.

Paul

[ric]

Originally Posted by JHagerson

It appears that the least performance degradation is to assign one LL test and one factoring assignment to each physical processor. You don't get full doubling, but it appears that the contention for the FPU is minimized.

Not the same everywhere: I've pretty much the same configuration (dual Xeon @ 3Ghz, HT enabled, 400MHz DDR2, Win2003 SP1), and I discovered that - due to the narrow bus of the Nocona Xeon - I achieve best results in assigning just a LL per physical unit. YMMV of course, but my advice is to give it a try under different conditions and choose what's best for your crunching box.

ric

[pcr]

Of course, the next step in the fun: is what Windows reports as "CPU 1" actually the second physical CPU or the second logical processor in CPU 0?

[JHagerson]

Originally Posted by xilman

Microsoft have been shipping and using multiprocessor-capable operating systems for years.

I looked at the Microsoft web site. They have developed some server-class offerings. Thank you for the pointer.

[xilman]

Originally Posted by JHagerson

I looked at the Microsoft web site. They have developed some server-class offerings. Thank you for the pointer.

You're welcome.

Did you find this one? http://www.microsoft.com/windowsserv...atacenter.mspx

It supports up to 64 cpus and a terabyte of RAM. Minimum requirement is 8 cpus.

While a 64-way box is far from being a supercomputer, it does constitute a reasonable powerful machine.


Paul