Prime95 version 25.7

[James Heinrich]

A few small (as in "easy" or "unimportant" is up to you) requests:

1) When LowMemWhileRunning is invoked, a message to that effect is printed in the Main Thread window (this is good). When PauseWhileRunning is invoked, only the worker window(s) get that message, not the Main window -- I'd prefer to see that kind of message in the main window (in the worker as well, if you want, but at least in the Main Thread window).

2) When PauseWhileRunning is triggered by ProgramX, a message to that effect is noted in the worker window. If ProgramY is started (which is also on the PauseWhileRunning list) while ProgramX is already running (and the worker is therefore paused), when you close ProgramX (but ProgramY is still running) the worker correctly does not resume, but it should print another message saying that it's now paused because ProgramY is running.

3) When starting workers, a check should be made for PauseWhileRunning and LowMemWhileRunning *before* an attempt to start is made -- I've seen workers start for several seconds (up to the PauseCheckInterval value) before noticing that ProgramX was running, then shifting into pause mode.

[Prime95]

Originally Posted by Axel1

is the server down at the moment? Get an unexpected CURL Lib Error?

These occur on occasion. Except for one user they appear to be intermittent.

Scott says the server will be down for maintenance for a short time tonight.

[James Heinrich]

Originally Posted by Prime95

No luck so far in fixing and memory bugs. Though I will compile 25.8 with the 3GB option for win32 and I guess more fox x64.

Since no 25.8 x64 build has appeared (yet), I'm starting to run P-1.stage2 (on M332203901) using the 32-bit version with 2500MB allocated, and so far it seems to be working. Although I have noticed that stage2 (at least with these abnormally-large candidates) is severely choked by memory... When running stage1 I had 4 threads running on my Q6600 (800MHz RAM) and I was getting reasonable usage out of it, around 92% CPU. Switch to stage2 and I was getting only 73% with 4 threads! 3 threads on P-1 and 1 thread on TF brought it up to 85% CPU, switching it to 2 threads on P-1 and 2 seperate TF threads gives me an overall 95% CPU usage. Obviously the stage2 is going to take a bit longer, but I'll at least get some other TF work done at the same time. I may have to borrow some DDR3 (ideally 1600MHz) and see how that improves my situation.

[Xyzzy]

On our laptop, TF (2 cores on 2 exponents) runs ~60°C and LL (2 cores on 1 exponent) runs ~70°C. We're not sure if that is interesting or not.

• Small feature request #1: Can we have an option to change console fonts?
• Small feature request #2: Is it possible to have a log viewer (results.txt & prime.log) built into the client?
• Small feature request #3: On dual core machines, a screen layout option with the top 1/3 of the console set for server communications and the bottom 2/3 of the console split in half for each core would be cool. We can set it up manually but a button to do it would be nice.

[garo]

Sending benchmark data to server
...
pnErrorResult=7
pnErrorDetail=parameter FFT4K: Invalid float value: '5.40571e-005'

I just sent a PM to George about this. It looks like this error has mutated into "invalid float precision" in v25.8. But I get this while reporting end dates on normal exponents.

[James Heinrich]

Originally Posted by garo

It looks like this error has mutated into "invalid float precision" in v25.8. But I get this while reporting end dates on normal exponents.

As I said a few messages up, v25.8.1 worked fine for me in fullbench mode, I got neither local nor server error messages.

[garo]

Right! I did not ackowledge that in my post above but did read that which is why I pointed out that it happened to me while communicating end dates to the server.

[James Heinrich]

Originally Posted by James Heinrich

I may have to borrow some DDR3 (ideally 1600MHz) and see how that improves my situation.

So, I did -- borrowed some nice DDR3 (4GB of Patriot Viper DDR3-1600MHz, 7-7-7-20), and since my motherboard (Gigabyte GA-P35C-DS3R) can support either DDR2 or DDR3 I can make direct comparisons on the same system. I ran a worker for stage2 P-1 on M332203901 with 1/2/3/4 threads on both DDR2@768MHz,5-5-5-15 and DDR3@1536,7-7-7-20 (odd RAM speeds are as close to rated RAM speeds as I can get running my Q6600 at 3.45GHz) and averaged the times over 3 sets of 100 iterations. Worker output looks generally like this, if it's of any relevance:

Worker starting
Setting affinity to run worker on any logical CPU.
Optimal P-1 factoring of M332203901 using up to 2500MB of memory.
Assuming no factors below 2^75 and 2 primality tests saved if a factor is found.
Optimal bounds are B1=3360000, B2=91560000
Chance of finding a factor is an estimated 6.91%
Using FFT length 20M, 2 threads
Setting affinity to run helper thread 1 on any logical CPU.
Available memory is 2498MB.
Using 2468MB of memory.  Processing 10 relative primes (20 of 480 already processed).
M332203901 stage 2 is 5.26% complete.

I expected the CPU usage to go up when more RAM bandwidth was available, but that was entirely not the case. Overall performance was better with DDR3 than DDR2 (by 7%, 17%, 18%, 23% for 1,2,3,4 threads respectively -- see attached graph), but CPU usage was identical between both sets of tests. Benchmark results show a similar performance gain with 4 threads (+22%), but performance increase drops to around 3% for 1 thread.
Benchmarks, and Excel file used to generate this graph, in this post.

[Jeff Gilchrist]

Originally Posted by James Heinrich

I expected the CPU usage to go up when more RAM bandwidth was available, but that was entirely not the case. Overall performance was better with DDR3 than DDR2 (by 7%, 17%, 18%, 23% for 1,2,3,4 threads respectively -- see attached graph), but CPU usage was identical between both sets of tests.

So I guess the memory bottleneck is still there so you don't see the CPU usage increase, but with DDR3 you can shove more through the gate in the same amount of time so it effectively increases performance.

[Prime95]

I'll bet waiting for memory does not reduce the OS's interpretation of your CPU usage. The OS probably counts the CPU as in use as long as a thread is bound to the processor.

[petrw1]

I have a

Software Version Windows,Prime95,v25.7,build 3
Model Intel(R) Pentium(R) 4 CPU 2.40GHz
Features Single core, 2 hyperthreaded CPUs, Prefetch,SSE,SSE2
Speed 2.394 GHz (2.465 GHz P4 effective equivalent)
L1/L2 Cache 8 / 512 KB
Computer Memory 511 MB configured usage 300 MB day / 300 MB night

In v24 it only used one CPU and Task Manager showed 50% used and 50% idle. I thought I could got a small benefit from the Helper-Thread concept of v25. I set the Worker as:

1 Worker
First Time LL Tests
Run on any CPU
Multithreading (CPUs to use): 2

CPU usage went to 100% so I assumed I must be getting better throughput even though the per iteration time seemed to be "about the same".

After a month or so I decided to try a suggestion from this forum to assign small work to the second CPU while letting the first alone do LL. So I went to:

2 Workers
Worker 1:
- First Time LL Tests
- Run on CPU 1
- Multithreading (CPUs to use): 1
Worker 2:
- TF
- Run on CPU 2
- Multithreading (CPUs to use): 1

BAD MOVE!!!

My iteration time on Worker 1 went from .063 to .109.
My time for 265 TF iterations was over 6 seconds.

When I selectively stopped Worker 2; Worker 1 went to .061 seconds.
When I stopped 1 and started 2; Worker 2 went to 4.5 seconds.
I tinkered with reversing the CPU assignments and also Smart Assignment.
Same results each time.

Hey, wait a minute:  Worker 1 is faster one 1 CPU than both?!?!?!?

So I went back to 1 worker but only using 1 CPU (Run on any CPU) and my iteration time stayed at .061 instead of .063 (about 3% faster!)

Can anyone explain why my Helper Thread made it worse?
Is this possible a case where v25.8 will help?