Prime95 30.8 (big P-1 changes, see post #551)

[kriesel]

Originally Posted by kriesel

Two outstanding issues as far as I know (haven't tested v30.8b17 yet)

1) Observed on Windows 7 Pro x64, dual Xeon E5-2670, prime95 V30.8b15, using start /Node 0 or 1, /affinity 0x5555, running two instances, intended as one each side of the QPI;
when a worker window assigns cores, the following message is produced repeatedly, with variety of hex values consisting of 3 or c at various offsets:
Error setting affinity to cpuset 0x000000c0: No error
(refer to attachment of https://mersenneforum.org/showpost.p...&postcount=731)
3 or c is 0011 or 1100.
Windows' numbering representation of the two logical cores of a x2 hyperthreaded physical core #0 is 0,1,
while Linux's is 0,n where n is number of physical hyperthreaded cores present in the system.
So it appears to me that prime95, a Windows application, may be using an inappropriate affinity mask for Windows. We don't usually want two prime95 compute threads running on the same physical core.

... except when using hyperthreading while performing TF, which we'd rather do on GPUs anyway for performance. I've been starting the dual instances with the equivalent of /node 0 and /node 1 of the following,

start /node 0 /affinity 0x5555 /d "C:\Users\ ... \prime95-x64" prime95.exe

It never occurred to me that prime95 would default to no hyperthreading for fft multiplies, and generally benchmark as better performance without using hyperthreading, yet insist on being able to set both hyperthreads of the same physical cores as available, even trying to override the user.

I should be using 0xFFFF, which allows 16 threads on 8 cores? And let the threads flop about between logical processors? Seems counter to documentation's guidance. From readme.txt:

Use hyperthreading
------------------

Except for trial factoring, which is best left for GPUs to do, hyperthreading often offers no performance
benefit while using more electricity.  You can try test if hyperthreading speeds up your worker windows by
selecting these options.

Or for that matter, "not recommended" built into the GUI dialog boxes.

I found in early experimentation that without the affinity mask included, the Windows 7 start /Node 1 command did not properly place the second instance on NUMA node 1, instead running both instances on the various hyperthreads of node 0 (CPU package 0), saturating all its logical processors, and leaving the second processor package idle, reducing performance.

To restate:

start /node 0 /affinity 0x5555 /d "C:\Users\ ... \prime95-x64" prime95.exe
start /node 1 /affinity 0x5555 /d "C:\Users\ ... \2\prime95-x64" prime95.exe

Good performance, but error messages;

without affinity masks,

start /node 0 /d "C:\Users\ ... \prime95-x64" prime95.exe
start /node 1 /d "C:\Users\ ... \2\prime95-x64" prime95.exe

No error messages, but both land on node 0, so half the performance.

Using affinity mask 0xFFFF in the start command seems to resolve the error messages, but the selection of hyperthreads is irregular in that case. Its hyperthread use is also not stable over time, switching between logical cores of a physical core somewhat.

[kriesel]

On the dual E5-2670 Win 7, again:

Node 1, using affinity mask 0xFFFF, is alternately using all 16 logical cores, rapidly changing between them. This node is doing P-1. Affinity related error messages remain, plus there's a new variant. The error message does not relate to proof storage optional directory, or temporary directory, because those fields are blank. The path specified in the start command is known to be valid since it was copied/pasted from Explorer, and is on the system's boot drive.

[Oct 2 15:55:30] Worker starting
[Oct 2 15:55:30] Setting affinity to run worker on CPU core #1
[Oct 2 15:55:30] Error setting affinity to cpuset 0x00000003: No such file or directory
[Oct 2 15:55:30] Optimal P-1 factoring of M116509073 using up to 49152MB of memory.
[Oct 2 15:55:30] Assuming no factors below 2^77 and 1.3 primality tests saved if a factor is found.
[Oct 2 15:55:30] Optimal bounds are B1=805000, B2=201211000
[Oct 2 15:55:30] Chance of finding a factor is an estimated 5.23%
[Oct 2 15:55:30] 
[Oct 2 15:55:32] Setting affinity to run helper thread 1 on CPU core #2
[Oct 2 15:55:32] Setting affinity to run helper thread 2 on CPU core #3
[Oct 2 15:55:32] Using AVX FFT length 6400K, Pass1=640, Pass2=10K, clm=2, 4 threads
[Oct 2 15:55:32] Setting affinity to run helper thread 3 on CPU core #4
[Oct 2 15:55:32] Error setting affinity to cpuset 0x0000000c: No error
[Oct 2 15:55:32] Error setting affinity to cpuset 0x00000030: No error
[Oct 2 15:55:32] Error setting affinity to cpuset 0x000000c0: No error
[Oct 2 15:55:33] Ignoring suggested B2 value, using B2=201653100 from the save file
[Oct 2 15:55:37] Available memory is 49089MB.
[Oct 2 15:55:39] Setting affinity to run helper thread 1 on CPU core #2
[Oct 2 15:55:39] Setting affinity to run helper thread 2 on CPU core #3
[Oct 2 15:55:39] Error setting affinity to cpuset 0x00000030: No error
[Oct 2 15:55:39] Switching to AVX FFT length 7M, Pass1=448, Pass2=16K, clm=4, 4 threads
[Oct 2 15:55:39] Estimated stage 2 vs. stage 1 runtime ratio: 0.890
[Oct 2 15:55:39] Error setting affinity to cpuset 0x0000000c: No error
[Oct 2 15:55:39] Setting affinity to run helper thread 3 on CPU core #4
[Oct 2 15:55:39] Error setting affinity to cpuset 0x000000c0: No error
[Oct 2 15:55:40] Using 49055MB of memory.  D: 1650, 200x687 polynomial multiplication.
[Oct 2 15:55:51] Setting affinity to run polymult helper thread on CPU core #2
[Oct 2 15:55:51] Setting affinity to run polymult helper thread on CPU core #3
[Oct 2 15:55:51] Setting affinity to run polymult helper thread on CPU core #4
[Oct 2 15:55:51] Error setting affinity to cpuset 0x0000000c: No error
[Oct 2 15:55:51] Error setting affinity to cpuset 0x00000030: No error
[Oct 2 15:55:51] Error setting affinity to cpuset 0x000000c0: No error
[Oct 2 15:58:41] Stage 2 init complete. 5235 transforms. Time: 186.005 sec.
[Oct 2 15:59:52] M116509073 stage 2 at B2=58167450 [16.98%]
[Oct 2 16:25:20] M116509073 stage 2 at B2=72373950 [25.20%].  Time: 1528.234 sec.
[Oct 2 16:51:05] M116509073 stage 2 at B2=86580450 [33.42%].  Time: 1544.864 sec.

[kriesel]

Same system as previous post:
A second instance, specified in start command to run on NUMA node 1, without affinity mask used, lands on node 0 instead. But no error messages. Note first instance's iteration time more than doubled, and NUMA node 1 is idle.

[storm5510]

Originally Posted by kriesel

Same system as previous post:
A second instance, specified in start command to run on NUMA node 1, without affinity mask used, lands on node 0 instead. But no error messages. Note first instance's iteration time more than doubled, and NUMA node 1 is idle.

NUMA: Non-Uniform Memory Access.

Off-topic: Why would the BIOS in a single CPU system have this enabled? It is on my old Xeon...

[bplenhart]

Are there any plans to update the FreeBSD version past 30.7b9?

[Prime95]

Originally Posted by bplenhart

Are there any plans to update the FreeBSD version past 30.7b9?

Try 30.9b1: https://www.dropbox.com/s/1ag6cxl8f5...64.tar.gz?dl=0

It actually will be easier to build 30.9b3 rather than a 30.8 version.

[tha]

In light of all these changes, the math page listed on the home page could use an update.

[kriesel]

After a stop and continue during stage 2 P-1, the output line

Estimated stage 2 vs. stage 1 runtime ratio: x.xxx

seems to consider only the remaining portion of stage 2, not full stage 2 run time.
So for an actual stage 2 / stage 1 ratio ~1, if stage 2 is 80% complete when resumed, the resumption will indicate a ratio of ~0.2.

[kriesel]

Somewhat similar to how gpuowl does things, please add an option in prime95 to log everything that appears in a worker window to a log file (open for append, write, close). Overnight, Windows Update restarted a system that was in the midst of a ~3.5 day 100Mdigit P-1 benchmarking run, and all prime95 worker window content (timing for stages & GCD) was lost. That content does not appear anywhere else (prime.log, results.txt, etc.).

[kriesel]

This combo reports various cpu speeds, with the value in the second line fluctuating around 1/4-1/3 of the actual.
I've seen 440 MHz down to 371 MHz while Task Manager shows 1500 MHz (slight turbo clock).
Clearing the CpuSpeed entry out of local.txt and restarting the program does not resolve the issue.
This really throws off the ETA computations.

Its 7210 sibling running v30.7b9 reports cpu speed correctly.

[pepi37]

Before two days my Intel I5 9600 ( never OC) but working with Prime95 few years 24/7 start to show warnings Confidence in final result is very poor. So what else to do then go and buy Ryzen5 3600 and restart sequence for test. And why I wrote all this: Beside warning I was surprised when new CPU give me exact res as my "old Intel". So for me Prime95 is best software I ever and will use in future.!