Prime95 version 28.9 / 28.10

[S485122]

As far as I understand it the OS "thinks" it is working with cores, not threads (that is another concept in the world of OS's.)
Then there is no such thing as a a real or physical core on one side and a virtual core on the other. A CPU has a number of cores, the OS "sees" that number of cores but in the case of a hyper-threading capable CPU the OS "sees" twice that number. You just have pairs of virtual cores sharing one physical core. The OS sees one hyper-threaded core as two identical cores apart from the core number assigned to each.

This means that a Windows affinity scramble of 02468A13579B is equivalent to 13579B02468A and a Linux affinity scramble of 0123456789AB is equivalent to 6789AB012345 . There might be a difference in the order in which the OS tries to use threads though.

Jacob

[pepi37]

Can you in next release add option to stop processing next candidate of K if K found PRP ( like LLR do)?

from LLR
StopOnPrimedK=1 : after successes with this k value,skip further pairs having the same k value

[Uncwilly]

I upgraded one of my machines (an i5 with 4 cores, AVX, but not AVX2) to the Win 64-bit version of 28.9 (from an older 64-bit version).

I had it running TFs on numbers in the 332M range. (Yes, I know, but I don't have any GPUs myself, but have idle CPUs).

On firing up 28.9 it moved all of the worktodo.txt entries into [Worker 1]. All well and good....
But, it was only using 2 of the 4 cores, until I monkeyed with the settings to give it helper threads. Shouldn't it have applied all cores to the new task if it moved all of the worktodo to that thread?

With all cores running on a single exponent, it looks like I am getting the same throughput as an identical machine running an older version, +/- 1%.

[Prime95]

Originally Posted by Uncwilly

I upgraded one of my machines (an i5 with 4 cores, AVX, but not AVX2) to the Win 64-bit version of 28.9 (from an older 64-bit version).

I had it running TFs on numbers in the 332M range. (Yes, I know, but I don't have any GPUs myself, but have idle CPUs).

On firing up 28.9 it moved all of the worktodo.txt entries into [Worker 1]. All well and good....
But, it was only using 2 of the 4 cores, until I monkeyed with the settings to give it helper threads. Shouldn't it have applied all cores to the new task if it moved all of the worktodo to that thread?

With all cores running on a single exponent, it looks like I am getting the same throughput as an identical machine running an older version, +/- 1%.

When you upgraded to 28.9, the number of workers and number of CPUs to use for each worker were left unchanged. Only on fresh installs will 28.9 set num workers to 1 with maximum multithreading.

If there are more worker sections than number of workers, this version does move worktodo lines accordingly.

The +/-1% is expected as improvements came from AVX2 instructions.

[ssateneth]

It seems that 28.9 added multithreading to lucas-lehmer prime testing compared to 28.7. Maybe this is undocumented? I didnt see this in whats new, and 28.7 didn't have this behavior last I checked. 1 single primality test is using 4 cores. I have a 5960x. Is it possible to increase the amount of threads used for 1 task up from 4? I want to try using 8 and 16 for the same task, if possible. I didn't see anything in undoc on manually setting helper threads, just some things about affinity scrambling.

[ixfd64]

Originally Posted by ssateneth

It seems that 28.9 added multithreading to lucas-lehmer prime testing compared to 28.7. Maybe this is undocumented? I didnt see this in whats new, and 28.7 didn't have this behavior last I checked. 1 single primality test is using 4 cores. I have a 5960x. Is it possible to increase the amount of threads used for 1 task up from 4? I want to try using 8 and 16 for the same task, if possible. I didn't see anything in undoc on manually setting helper threads, just some things about affinity scrambling.

From the changelog:

8)  On initial install, the default settings for number of workers will be set to
    the number of cores / 4 with multithreading turned on.

[TObject]

ssateneth, the threading settings are under "Test - Worker Windows..." menu. Once you wrap your hand around how that works, you should be able to set what you desire, as long as there are enough cores available.

[Madpoo]

Originally Posted by ssateneth

It seems that 28.9 added multithreading to lucas-lehmer prime testing compared to 28.7. Maybe this is undocumented? I didnt see this in whats new, and 28.7 didn't have this behavior last I checked. 1 single primality test is using 4 cores. I have a 5960x. Is it possible to increase the amount of threads used for 1 task up from 4? I want to try using 8 and 16 for the same task, if possible. I didn't see anything in undoc on manually setting helper threads, just some things about affinity scrambling.

It's been available for a while but it was probably poorly understood by new users. The new version changes some defaults to make testing more efficient on memory bandwidth limited systems (which is pretty much all of the current testing going on now, at the current FFT sizes).

Set it to whatever, but don't have a worker use more threads than you have CPU cores (or more workers * threads than actual CPU cores, in general).

[M0CZY]

When using the Linux version of mprime 28.9, if I don't want new tasks sent, the terminal menu doesn't allow me to set 'days of work' to 0.
Instead I am forced to manually edit the file prime.txt to read DaysOfWork=0

Is this issue something that can be rectified in the next version?

[Dubslow]

NoMoreWork=1

[M0CZY]

NoMoreWork=1

Yes, but that’s just doing what I'm already forced to do, but in a different way.

What I was requesting was if it was possible that a modification could be made in a future version to allow the terminal menu 14 to accept 'Days of work to queue up' to be set to 0, which it currently doesn't.