v25.7 not using much cpu

[dan3ny]

I just upgraded from v24 to 25.7, build 3.

Running an Intel T2500 @ 2Ghz, 2 core WinXPSP2

Upgraded from 2 instances of p95 to one, with 2 worker threads. Where my old instances used to keep the CPU pegged near 100% between the 2 of them, the new version isn't using very much at all. I found that I had a Throttle command in my prime95.txt from before, but I removed it.

Why isn't p95 using my CPU??

[Uncwilly]

What numbers (range) and type of work are you doing?
Are both workers running on 1 number, or are each doing their own?
I found that 2 workers on factoring a single number has ~55% the throughput of 2 workers factoring 2 different numbers.

[Oleg V.Cat]

Originally Posted by Uncwilly

Are both workers running on 1 number, or are each doing their own?
I found that 2 workers on factoring a single number has ~55% the throughput of 2 workers factoring 2 different numbers.

I can say, that sometimes "Smart affinity" is very stupid. I do not have problems on single CPU double Core, bur, for example on 4*XeonMP if I put "4 workers 2 threads each" - two of workers go on #CPU 0,4 and two - on #CPU 1,5, so, I have only 50% cpu used for 8 threads.

On other box I have 2*L5420 (Quad core), and i have same throughput for:
4*V24, with affinity 0,3,4,6 and 2*V24+1*25(2 threads), making double check, but if I use any other core combination (not 0,3,4,6) - througput is dropped by ~30% at least :(

[dan3ny]

Originally Posted by Uncwilly

What numbers (range) and type of work are you doing?
Are both workers running on 1 number, or are each doing their own?
I found that 2 workers on factoring a single number has ~55% the throughput of 2 workers factoring 2 different numbers.

Factoring 2 numbers in the 4X,XXX,XXX range, each doing their own.

[dan3ny]

Hmm. It looks like the Throttle command came back or I didn't kill it completely somehow. Removed it again, and now things seem fine.

[Prime95]

Originally Posted by Oleg V.Cat

I can say, that sometimes "Smart affinity" is very stupid. I do not have problems on single CPU double Core, bur, for example on 4*XeonMP if I put "4 workers 2 threads each" - two of workers go on #CPU 0,4 and two - on #CPU 1,5, so, I have only 50% cpu used for 8 threads.

On other box I have 2*L5420 (Quad core), and i have same throughput for:
4*V24, with affinity 0,3,4,6 and 2*V24+1*25(2 threads), making double check, but if I use any other core combination (not 0,3,4,6) - througput is dropped by ~30% at least :(

You lost me completely! I'm guessing you think the 4*XeonMP should be CPU 0,1 and CPU 2,3, etc. I can't figure out what you are doing on the 2*L5420 and what you think it should be doing.

[Oleg V.Cat]

Originally Posted by Prime95

You lost me completely! I'm guessing you think the 4*XeonMP should be CPU 0,1 and CPU 2,3, etc. I can't figure out what you are doing on the 2*L5420 and what you think it should be doing.

No ;). 4*XeonMP shoud be CPU 0,2,4,6 (or 1,3,5,7) for best performance and lower CPU usage (ThreadsPerTest=1). In this case I have ~0.048s per iteration (all timings are for LL with FFT1280K) with 50% CPU. If 8 workers are running, I have ~0.085s per iteration, okay, it's a bit better that 0.048*2, but no so much, and i prefer to use only 4 workers.

If I try to set (ThreadsPerTest=2) - tu use all 8 logical CPU for 4 workers I need to set affinity to 0,1,4,5 (prime95 then select 2,3,6,7 for helper threads) - in this case I have 100% CPU and, approx, ~0.047s per iteration, so, no advantage from helper thread. I want to try set main thread to CPU0 and helper thread to CPU1, but can't see, how.

On 2*L5420 picture is more complex :). Again, i do not know why, but if I run just one worker with ThreadsPerTest=1 - i have ~0.028 per iteration. If one (or more) workers added - throughput go down, but very differently, depending on affinity. 4 worker thread combinations can be down to ~0.055s in "bad" combination, or can be ~0.030 for "good". At this moment I have 3 workers with foillowing params:

Worker No. Threads CPU timing
1 1 0 0.029
2 2 3,4 0.016
3 3 6 0.031

Any other combination give ~0.042 per iteration for, at least, 2 workers. I do not know, why, "it's magic" ;)

[Freightyard]

The Xeon 5400 Series Quad-Core CPUs are all twin dual-core dies. Each die (or pair of cores) contains 6 MB of the total 12 MB L2 cache:

http://download.intel.com/design/xeo...pdt/318585.pdf

Quad-core desktop CPUs are similar.

Hence, threads on cores 0 and 1 compete for the same cache, but do not compete for the cache shared by cores 2 and 3.

Threads running (as you state) on cores 0/3/4/6 would all have unique cache and would not compete. But 0/2/4/6 should equally not compete.

The unknown is your OS and other processes, which may be cache- or CPU- intensive. Without knowing all these unknowns it certainly could seem like "magic". You may be able to set affinity for some of the system threads to increase performance.

[Prime95]

Originally Posted by Oleg V.Cat

I can say, that sometimes "Smart affinity" is very stupid. I do not have problems on single CPU double Core, bur, for example on 4*XeonMP if I put "4 workers 2 threads each" - two of workers go on #CPU 0,4 and two - on #CPU 1,5, so, I have only 50% cpu used for 8 threads.(

Please try this again on 25.8 when it is available. BTW, all documentation I've read says that 0,1,2,3 are your "real" CPUs and 4,5,6,7 are the matching hyper-threaded CPUs. Thus 25.8 should put your 4 workers on CPUs 0,5 and 1,5 and 2,6 and 3,7.

[Meikel]

Hi George,

then I can, by trial, prove your documentation wrong - at least on Win Vista 64Bit and Core i7 940, and I agree with Oleg V.Cat. That's what I did:

First, tell Prime95 to run on any cpu. Then start an LL-Test with 4 threads.
Use Task Manager to bind Prime95 manually to CPUs.

Binding to CPUs 0,1,2,3: Best per iteration time: 11.435ms
Binding to CPUs 0,2,4,6: Best per iteration time: 6.833ms

Seems quite obvious to me, no? CPUs 0,2,4,6 are "real". Definitively. Maybe Intel uses a different numbering scheme than Microsoft? Whatever: The measurements don't lie.

IMHO best performance for a Core i7 could be reached by assigning 4 LL tests to it, starting two threads for each LL test and assigning the first test to the first real CPU-HT-pair (0,1), the second to (2,3), and so on. I also agree with Oleg on that.

So, I'm pretty convinced that your proposed assignment strategy for 25.8 would totally NOT work well :-)

[Prime95]

Originally Posted by Meikel

So, I'm pretty convinced that your proposed assignment strategy for 25.8 would totally NOT work well :-)

Alright, we'll go with 0,1 and 2,3 and 4,5 and 6,7.