Memory Thrashing

[storm5510]

Originally Posted by science_man_88

from Wikipedia:
...The term is also used for various similar phenomena, particularly movement between other levels of the memory hierarchy, where a process progresses slowly because significant time is being spent acquiring resources.

This is what I am referring to. In this case, it does not progress slowly, it pauses, then resumes normally.

[science_man_88]

Originally Posted by storm5510

This is what I am referring to. In this case, it does not progress slowly, it pauses, then resumes normally.

do you know the cache sizes etc. that might help if it's a memory thing, are you running anything else, and if so does it have a pause while running for it ?

[Madpoo]

Originally Posted by storm5510

This is what I am referring to. In this case, it does not progress slowly, it pauses, then resumes normally.

If it matters, I was confused by the term "memory thrashing" and initially I thought you meant just on the memory bus to RAM, but then I wasn't sure if your system was using too much memory and you meant it was paging to disk (I'd call that swapfile thrashing though).

As long as we're ruling out pagefile thrashing and you just meant the RAM itself...

There are in fact some weird memory effects I've seen with Prime95, specifically with multiple cores in a single worker. I don't know about the factoring algos though, I've only seen it with LL tests.

With large exponents (really, large FFT sizes), it's not too bad to have 14+ cores in a single worker, and others have done 20+. When I do that, looking at the CPU shows each core is close to 100% usage so there's not too much "waste".

The weirdness comes when doing small FFT sizes, and I mean like 1MB or smaller. In those cases, each core is using far less than it's full potential, and I think as George explained it, those smaller sizes show the disparity between CPU and memory speed more clearly. They spend more time shuffling memory around between each "chunk" of work the cores are doing, so they idle quite a bit.

In extreme cases, like when I was doing all exponents below 1e6 in size, it was horrible, like each core in a 10-core worker was maybe doing 20%. It really made no sense that way and was better to actually run 10 single-core workers.

So... you may have a situation where CPU is finishing work faster than memory can do it's thing.

[storm5510]

Originally Posted by science_man_88

do you know the cache sizes etc. that might help if it's a memory thing, are you running anything else, and if so does it have a pause while running for it ?

No, not running anything else. Take a look at the image below.

Originally Posted by Madpoo

So... you may have a situation where CPU is finishing work faster than memory can do it's thing.

I never thought about this. It may be possible. They are DDR4-2400's I also considered that it may have something to do with the old GTX-480; buffering and so on. I am not sure how that works.

I got the term "thrashing" from George's readme.txt document which he includes with Prime95:

Be conservative. It is better to set the memory too low than too high. Setting the value too high can cause thrashing which slows down all programs. Remember, the program will only use the extra memory in stage 2 of P-1 factoring

This i7 has four cores and eight threads. It was suggested that I use one worker with four cores. It worked out well after a bit of adjusting. I can look at the utilization in the task manager. Sometimes, the four cores are maxed, but most of the time, not. It varies.