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

[kriesel]

Rule out intermittent page/swap slowdown first.

[storm5510]

Originally Posted by petrw1

I have one PC i5-6500; 16 GB RAM; 30.8b17.
Three times while I was on vacation it "seemed" to crash.
1. Late January: after 50 hours of no results reported it just started again. No one was near the PC during that time.
2. Early February: same scenario but this time 20 hours.
3. April 8th: Ditto; this time I came home 4 days later and it was still NOT making progress.
---- The screen showed daily communications to Primenet but not progress on the assignment.
---- Task Manager showed it using 2-5% CPU and nothing else using any more.
---- After a few restarts Prime95 CPU is now at 98% but it is still running very SLOOOOW. As in 12 seconds for 9 iterations (this is NOT a typo).
---- About an hour later all is well again (for how long???) ... 0.004 seconds for 9 iterations
---- CPU-Z shows CPU speed mostly at 3200 but occasionally 800, 1600 or 2400

I suspect bad hardware ... any suggestions?

As a start, I would upgrade it to 30.10 B? whatever it may be. Then, run a torture test. It this fails then test the RAM with something like Memtest86. Also, later model CPU's will throttle down if they get too hot. They lower their clock speed. There may be a CPU cooling problem.

[petrw1]

[Main thread Apr 12 10:37] Mersenne number primality test program version 30.8
[Main thread Apr 12 10:37] Optimizing for CPU architecture: Core i3/i5/i7, L2 cache size: 4x256 KB, L3 cache size: 6 MB
[Main thread Apr 12 10:37] Starting workers.
[Work thread Apr 12 10:37] Worker starting
[Work thread Apr 12 10:37] Setting affinity to run worker on CPU core #1
[Work thread Apr 12 10:37] 
[Work thread Apr 12 10:37] P-1 on M3168679 with B1=9500000, B2=TBD
[Work thread Apr 12 10:37] Using AVX FFT length 160K, Pass1=640, Pass2=256, clm=2
[Work thread Apr 12 10:37] M3168679 stage 1 is 15.45% complete.
[Main thread Apr 12 11:07] In write_giant, unexpected write failure len = 98352
[Main thread Apr 12 11:07] In write_giant, unexpected write failure len = 99022

[storm5510]

Originally Posted by petrw1

[Main thread Apr 12 10:37] Mersenne number primality test program version 30.8
[Main thread Apr 12 10:37] Optimizing for CPU architecture: Core i3/i5/i7, L2 cache size: 4x256 KB, L3 cache size: 6 MB
[Main thread Apr 12 10:37] Starting workers.
[Work thread Apr 12 10:37] Worker starting
[Work thread Apr 12 10:37] Setting affinity to run worker on CPU core #1
[Work thread Apr 12 10:37] 
[Work thread Apr 12 10:37] P-1 on M3168679 with B1=9500000, B2=TBD
[Work thread Apr 12 10:37] Using AVX FFT length 160K, Pass1=640, Pass2=256, clm=2
[Work thread Apr 12 10:37] M3168679 stage 1 is 15.45% complete.
[Main thread Apr 12 11:07] In write_giant, unexpected write failure len = 98352
[Main thread Apr 12 11:07] In write_giant, unexpected write failure len = 99022

There should be something in local.txt which looks like this:

WorkerThreads=1
CoresPerTest=4

From what I see above, you're not using your CPU very well. I looked up the specs. It's four-cores, four-threads. I have an i5-3570 in an older workstation and my suggestion works fine with it.

[petrw1]

Originally Posted by storm5510

There should be something in local.txt which looks like this:

WorkerThreads=1
CoresPerTest=4

From what I see above, you're not using your CPU very well. I looked up the specs. It's four-cores, four-threads. I have an i5-3570 in an older workstation and my suggestion works fine with it.

I'm a little (lot?) OCD.
In order to squeeze every bit out of my PC I have determined that running 4 cores/1 worker produces more throughput that 4/1....by about 20-30%
But I don't want to have every assignment reported twice (Stage 1 and Stage 2).
So my setup is MaxHighMemWorkers=1 with max RAM allocated.
At any point in time 3 cores run Stage 1 and only 1 runs Stage 2.

However running this way after a week or two Stage 1 gets way ahead and I end of with the cores spending a bit of time churning through every assignment until 3 of the 4 cores finds 1 that still needs Stage 1.

Therefore, every couple weeks I Stop; Change to 4/1 and stack up the Stage 1 complete exponents for it.

[petrw1]

Originally Posted by petrw1

[Main thread Apr 12 10:37] Mersenne number primality test program version 30.8
[Main thread Apr 12 10:37] Optimizing for CPU architecture: Core i3/i5/i7, L2 cache size: 4x256 KB, L3 cache size: 6 MB
[Main thread Apr 12 10:37] Starting workers.
[Work thread Apr 12 10:37] Worker starting
[Work thread Apr 12 10:37] Setting affinity to run worker on CPU core #1
[Work thread Apr 12 10:37] 
[Work thread Apr 12 10:37] P-1 on M3168679 with B1=9500000, B2=TBD
[Work thread Apr 12 10:37] Using AVX FFT length 160K, Pass1=640, Pass2=256, clm=2
[Work thread Apr 12 10:37] M3168679 stage 1 is 15.45% complete.
[Main thread Apr 12 11:07] In write_giant, unexpected write failure len = 98352
[Main thread Apr 12 11:07] In write_giant, unexpected write failure len = 99022

Seems to have stopped occurring.
I cleaned up the disk; it was under 100K free.

[Mark Rose]

Originally Posted by petrw1

I'm a little (lot?) OCD.
In order to squeeze every bit out of my PC I have determined that running 4 cores/1 worker produces more throughput that 4/1....by about 20-30%

With my i5-6600's, the crossover point where doing 4 workers is better is at 6144 FFT. A 5120 FFT prefers 4 threads and a single worker. At a 1024 FFT, a single worker is 14% faster. I haven't benchmarked lower. This FFT transition applies between core frequencies of 3.4 to 3.6 all core.

The threshold lowers a bit when I disable turboboost and cap the frequency at 3.3 GHz. I lose very little performance capping the frequency. My systems don't support unclocking or I'd clock lower. These CPUs are handicapped by the max supported DDR4-2133 memory speed, even with dual ranks per channel.

[slandrum]

Originally Posted by petrw1

Seems to have stopped occurring.
I cleaned up the disk; it was under 100K free.

Modern systems are not happy if they don't have a lot of the disk space free. If you have read-only media, that can be full, but if you have media that you do a lot of reading and writing to, things start to slow down when you get over about 80% full in my experience.

[storm5510]

Originally Posted by petrw1

Seems to have stopped occurring.
I cleaned up the disk; it was under 100K free.

I am surprised your OS didn't give you a low space warning. Windows does and I would think Linux should as well.

[petrw1]

Originally Posted by storm5510

I am surprised your OS didn't give you a low space warning. Windows does and I would think Linux should as well.

I did get a warning but assumed 100K was enough to run until I could cleanup more.

PS I ruled out heat since it runs weeks or months before it slows; and it slows to almost nothing; not just a small slow down...then the slow down lasts days, way too long to cool then quickly goes back to 100% for weeks again.

[alpertron]

From a puzzle in https://www.mersenneforum.org/showth...t=28312&page=7, I ran PRP of M281903623 / known 3 prime factors using Prime95 version 30.8 build 17.

After 70 days, Prime95 finished this task and uploaded the status, but the proof was not uploaded. I tried to upload it manually, but the server responded the following: