Prime95 Displays -1 ms/iter

[AI Complex]

I'm having a problem where Prime95 isn't properly showing the ms/iter and therefore isn't giving me proper estimated completion dates. I will attach a screenshot.

Please help me with this annoying bug.

Sincerely,
AI Complex

[ET_]

Originally Posted by AI Complex

I'm having a problem where Prime95 isn't properly showing the ms/iter and therefore isn't giving me proper estimated completion dates. I will attach a screenshot.

Please help me with this annoying bug.

Sincerely,
AI Complex

To correctly troubleshoot the issue, we need to know:

- What computer is the software running on (CPU, clock frequency, eventual overclock)
- The Memory= key on your local.txt file
- For how many hours/day the software is scheduled to run
- The version of Prime95 (menu help/about)

[Xyzzy]

http://www.mersenneforum.org/showthread.php?t=22845

[AI Complex]

Originally Posted by ET_

To correctly troubleshoot the issue, we need to know:

- What computer is the software running on (CPU, clock frequency, eventual overclock)

Attached is My PC's CPU-Z Report

Originally Posted by ET_

- The Memory= key on your local.txt file

Memory=1024 during 5:30-23:00 else 2048

Originally Posted by ET_

- For how many hours/day the software is scheduled to run
- The version of Prime95 (menu help/about)

24 hours per day (but the computer isn't always on, seeing as it is a laptop).
I'm running:

Windows64,Prime95,v29.4,build 8

[sdbardwick]

I have a feeling that HWLoc doesn't like AMD.
Version 29.4 doesn't like my Opteron 4280 box. Does silly CPU assignments - reports 8 hyperthreaded cores and assigns all 4 threads to 1 core (4 workers so 4 total cores used to handle 16 threads with 12 cores idle) and gives the same nonsensical iteration times.

[Prime95]

Originally Posted by sdbardwick

I have a feeling that HWLoc doesn't like AMD.
Version 29.4 doesn't like my Opteron 4280 box. Does silly CPU assignments - reports 8 hyperthreaded cores and assigns all 4 threads to 1 core (4 workers so 4 total cores used to handle 16 threads with 12 cores idle) and gives the same nonsensical iteration times.

I don't see how hwloc could affect the iteration timing code.

To see hwloc's analysis of your machine, do a quick benchmark and look in results.txt. If you think the analysis is wrong, post the results here and/or try to find a newer version of hwloc.dll

[sdbardwick]

I'm going to do a bare-metal reinstall of Win 7 x64 and Prime95 (v28 and v29 in separate directories) before I do any more investigation.

I ran into the issues with v29 a while back and gave up (simply went back to 28.5) because after a few attempts at getting core affinity right, I really needed my space heater. Now enough time has passed that I've forgotten the intricate details.

I'm more worried about not being able to get rational core assignments rather than properly displaying timings. Even categorizing it as hyperthreaded is annoying, as treating it like an Intel hyperthreaded system isn't optimal, based on my brief testing at the time.

BTW, this is the dual-socket development box resurrected for winter heater duty; I wonder if Prime95 isn't expecting the results hwloc provides with a dual socket NUMA AMD system.

Benchmark HW info from back when I was investigating:

AMD Opteron(tm) Processor 4280                 
CPU speed: 3100.14 MHz, 8 hyperthreaded cores
CPU features: 3DNow! Prefetch, SSE, SSE2, SSE4, AVX, FMA
L1 cache size: 16 KB
L2 cache size: 2 MB, L3 cache size: 6 MB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
L1 TLBS: 32
L2 TLBS: 1024
Machine topology as determined by hwloc library:
 Machine#0 (total=13152436KB, Backend=Windows, hwlocVersion=1.11.9, ProcessName=prime95.exe)
  NUMANode#0 (local=6544308KB, total=6544308KB)
    Package#0 (CPUVendor=AuthenticAMD, CPUFamilyNumber=21, CPUModelNumber=1, CPUModel="AMD Opteron(tm) Processor 4280                 ", CPUStepping=2)
      L3 (size=6144KB, linesize=64, ways=1, Inclusive=0)
        Core (cpuset: 0x00000003)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#0 (cpuset: 0x00000001)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#1 (cpuset: 0x00000002)
        Core (cpuset: 0x0000000c)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#2 (cpuset: 0x00000004)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#3 (cpuset: 0x00000008)
        Core (cpuset: 0x00000030)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#4 (cpuset: 0x00000010)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#5 (cpuset: 0x00000020)
        Core (cpuset: 0x000000c0)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#6 (cpuset: 0x00000040)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#7 (cpuset: 0x00000080)
  NUMANode#1 (local=6608128KB, total=6608128KB)
    Package#1 (CPUVendor=AuthenticAMD, CPUFamilyNumber=21, CPUModelNumber=1, CPUModel="AMD Opteron(tm) Processor 4280                 ", CPUStepping=2)
      L3 (size=6144KB, linesize=64, ways=1, Inclusive=0)
        Core (cpuset: 0x00000300)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#8 (cpuset: 0x00000100)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#9 (cpuset: 0x00000200)
        Core (cpuset: 0x00000c00)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#10 (cpuset: 0x00000400)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#11 (cpuset: 0x00000800)
        Core (cpuset: 0x00003000)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#12 (cpuset: 0x00001000)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#13 (cpuset: 0x00002000)
        Core (cpuset: 0x0000c000)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#14 (cpuset: 0x00004000)
          L2 (size=2048KB, linesize=64, ways=16, Inclusive=0)
            L1d (size=16KB, linesize=64, ways=4, Inclusive=0)
              PU#15 (cpuset: 0x00008000)
Prime95 64-bit version 29.4, RdtscTiming=1

[ATH]

My Intel 5960X does not like hwloc either, you can try and disable it with:
EnableSetAffinity=0
in prime.txt. With that setting I can use 29.x on my system.

[Prime95]

Originally Posted by sdbardwick

I'm going to do a bare-metal reinstall of Win 7 x64 and Prime95 (v28 and v29 in separate directories) before I do any more investigation.

I ran into the issues with v29 a while back and gave up (simply went back to 28.5) because after a few attempts at getting core affinity right, I really needed my space heater. Now enough time has passed that I've forgotten the intricate details.

I'm more worried about not being able to get rational core assignments rather than properly displaying timings. Even categorizing it as hyperthreaded is annoying, as treating it like an Intel hyperthreaded system isn't optimal, based on my brief testing at the time.

Prime95 uses a blend of old code and hwloc to identify the hardware. The hyperthreading identification is due to my old code, not hwloc. There is a local.txt setting to correct this (CpuNumHyperthreads or something similar). On my new machine the cache size is misidentified because of this old code -- I'll fix this next release to extract the info from hwloc instead.

The setting of affinity is done through hwloc using its CPU numbering scheme.

BTW, the latest hwloc is here: https://www.open-mpi.org/software/hwloc/v1.11/

[ixfd64]

I imagine it's related to this issue: http://mersenneforum.org/showpost.ph...&postcount=225