Worker Windows setting

[drew]

Hi,

For years, I've set my worker window settings to use Prime95 single-threaded by using 4 worker windows (1 per physical CPU). This optimizes the throughput, as far as I can tell.

For the first time I looked at the GIMPS assignment rules closely and realized these rules are set up in a way that prioritizes completing assignments quickly, giving faster workers smaller exponents.

This has the odd effect that by splitting up the work, which increases the duration of each test, I'll get assigned larger exponents, thereby taking even longer.

Should I be allocating the worker windows differently with the goal of completing smaller exponents sooner rather than optimizing the overall throughput?

[petrw1]

Originally Posted by drew

Hi,

For years, I've set my worker window settings to use Prime95 single-threaded by using 4 worker windows (1 per physical CPU). This optimizes the throughput, as far as I can tell.

For the first time I looked at the GIMPS assignment rules closely and realized these rules are set up in a way that prioritizes completing assignments quickly, giving faster workers smaller exponents.

This has the odd effect that by splitting up the work, which increases the duration of each test, I'll get assigned larger exponents, thereby taking even longer.

Should I be allocating the worker windows differently with the goal of completing smaller exponents sooner rather than optimizing the overall throughput?

My experience would suggest that changing to 2 workers with 2 cores each "should not" reduce overall thruput. That is, if running 4 workers each assignment takes 30 days then with 2 workers with 2 cores each assignment should be close to 15 days; maybe 16 days.

However extending to 1 worker with 4 cores will likely reduce overall thruput; I'll hazard a guess of 20 or 30%. Each assignment would not drop to 7.5 or 8 days but more likely 10. It depends a lot on amount or RAM; speed of RAM and the internal BUS/Cache etc. hocus-pocus (that means I'm a novice on this stuff).

That all said your best friend is simply to run a benchmark and it will tell you exactly.

[kriesel]

Don't guess. Benchmark. It depends on the processor, cache size, and the fft size, whether 1, 2, 4, etc cores/worker are maximal throughput overall. See the attachments at posts 4 and 5 of https://www.mersenneforum.org/showthread.php?t=23900
As exponents get larger, primality tests get longer, and more cores/worker may be needed to avoid expiration before assignment completion. This is especially the case with older slower multicore processors.

[Uncwilly]

The push for faster turn over with regards to the assignment rules is primarily about the trailing edge. If you are in Cat 2 or 3, you should not be having a problem finishing the exponents using 1 core per worker. Cat 0 and 1 will require 2 cores. (Provided you keep your Days of work to queue low.)

[drew]

Originally Posted by Uncwilly

The push for faster turn over with regards to the assignment rules is primarily about the trailing edge. If you are in Cat 2 or 3, you should not be having a problem finishing the exponents using 1 core per worker. Cat 0 and 1 will require 2 cores. (Provided you keep your Days of work to queue low.)

Ok thanks. I get the point about trying to capture the trailing edge. I'm not worried about being able to complete the work in time, I was just trying to understand what's in the best interest for GIMPS.

[kladner]

Originally Posted by petrw1

My experience would suggest that changing to 2 workers with 2 cores each "should not" reduce overall thruput. That is, if running 4 workers each assignment takes 30 days then with 2 workers with 2 cores each assignment should be close to 15 days; maybe 16 days.

However extending to 1 worker with 4 cores will likely reduce overall thruput; I'll hazard a guess of 20 or 30%. Each assignment would not drop to 7.5 or 8 days but more likely 10. It depends a lot on amount or RAM; speed of RAM and the internal BUS/Cache etc. hocus-pocus (that means I'm a novice on this stuff).

That all said your best friend is simply to run a benchmark and it will tell you exactly.

I'm not sure this is true. Would contrary benchmarks from a one worker, four core setup help cast light on the subject?

"Scuze me while I whip this out!"

(See attached.) I just stopped my normal DC work and ran this short-form benchmark. The CPU frequency of 4000MHz given is incorrect. The CPU runs 4 cores locked at 4300MHz. The BIOS shows this (4000MHz) at startup, as well. But a short test run of the benchmark with HWINFO running showed the CPU at 4300. The same is true with regular DCLL work.

[kriesel]

Originally Posted by Uncwilly

The push for faster turn over with regards to the assignment rules is primarily about the trailing edge. If you are in Cat 2 or 3, you should not be having a problem finishing the exponents using 1 core per worker. Cat 0 and 1 will require 2 cores. (Provided you keep your Days of work to queue low.)

It depends a great deal on how slow the cores are.

[petrw1]

Originally Posted by kladner

I'm not sure this is true. Would contrary benchmarks from a one worker, four core setup help cast light on the subject?

"Scuze me while I whip this out!"

(See attached.) I just stopped my normal DC work and ran this short-form benchmark. The CPU frequency of 4000MHz given is incorrect. The CPU runs 4 cores locked at 4300MHz. The BIOS shows this at startup, as well. But a short test run of the benchmark with HWINFO running showed the CPU at 4300. The same is true with regular DCLL work.

Hmmm as I read it I stand corrected.
Geez usually my 1 mistake of the year takes longer. LOL