CPU Worker Windows

[storm5510]

With this new i7-7700 CPU, Prime95 is only allowing one worker window spread across four threads. The other four threads are not being used. When I installed Prime95, I only included the application file itself. No local.txt or prime.txt. I didn't want to cause problems by using ones from a totally different system. Prime95 created new ones.

I browsed undoc.txt and added the three lines below into local.txt:

NumCPUs=4
CpuNumHyperthreads=1
CpuSpeed=3978

With this addition, I now have one worker and three helpers. Four other threads remain unused.

Setting affinity to run worker on CPU core #0.
Setting affinity to run helper thread 1 on CPU core #1
Setting affinity to run helper thread 2 on CPU core #2
Setting affinity to run helper thread 3 on CPU core #3

The only other addition I made is to add 'CumulativeTining=1" to prime.txt.

I'm not sure what else to do. What I was hoping for was four workers and four helpers. I must have really fudged something! Thoughts and ideas, please...

Thanks!



Note: Windows SmartScreen blocks Prime95 v29.2, Build 2, from running on Windows 10 Pro x64, at least, for me. This is a clean install so it may not have the updates it needs yet.

[thyw]

I think you should set CpuNumHyperthreads to 2, since i7-7700 is a hyperthreaded cpu.
(Also, P95 should've figured out these settings automatically on the first startup. By including these information you're overriding the p95 CPU detection.
"If the program did not correctly figure out your CPU information, you can override the info in local.txt:")

Edit: in sub v29!, looks like
To create 4 worker window (each with 1 main and 1 helper thread, logically on separeted cores per worker)
you would want to write

WorkerThreads=4
ThreadsPerTest=2 //this is global, you could inlude this in each "worker block" for different values
[Worker #1]
Affinity=0
[Worker #2]
Affinity=2
[Worker #3]
Affinity=4
[Worker #4]
Affinity=6

PS I'm not sure about these:
Wouldn't many (4) P-1 (really memory intensive) workers be strongly bottlenecked by memory?
Does HT benefit P-1?
In the new undoc (v29) Affinity you have main and aux (helper?) threads split accross multiple CPUS, how should it be used?
Also is 8 GB RAM enough for 4 p-1 job? (you can include MaxHighMemWorkers=n, but it means waiting)

[GP2]

Originally Posted by thyw

WorkerThreads=4
ThreadsPerTest=2 //this is global, you could inlude this in each "worker block" for

Version 29 uses "CoresPerTest" rather than "ThreadsPerTest" as previous versions did. It gets automatically converted when you run it, however.

[Prime95]

You'll find that using the hyperthreads will make your CPU run hotter and *reduce* your throughput. The benchmark you posted shows this.

In v29, the worker windows menu choice lets you choose the number of workers and how many CPU cores each worker will use. There is also a checkbox if you want LL tests to use hyperthreading on those cores (not recommended).

The menus should provide access to prime95's important features. You only need to dive into undoc.txt if you want to do strange stuff.

[storm5510]

With a little experimentation, I now have it running two workers on eight threads. I was not sure how far I could go with this...

CoresPerTest=4
WorkerThreads=2

I don't know if this was a correct way, but it 'appears' to function properly. Core temps are running in the mid '70's. The PSU is pulling 200W from the service. If it goes higher, my APC unit starts to whistle and the light changes to yellow.

Below are two snip from the task manager.

[Mark Rose]

Four threads will fully utilize the CPU though. You're making more heat for no benefit.

[storm5510]

Originally Posted by Mark Rose

Four threads will fully utilize the CPU though. You're making more heat for no benefit.

I need to start looking at this by physical cores and not logical. So I changed the settings:

CoresPerTest=2
WorkerThreads=2

In the new snip below, the utilization is 100%, at the top-right. Below the graphs it is roughly half. Four threads are idling, but not cores. Temps are in the mid 60;s. The throughput from Prime95 is virtually unchanged.

[Dubslow]

Originally Posted by storm5510

I need to start looking at this by physical cores and not logical.

Good way to think about Prime95.

Originally Posted by storm5510

Below the graphs it is roughly half. Four threads are idling, but not cores.

OS Task Managers don't generally tend to consider the effect of hyperthreads on actual silicon usage.

[Mark Rose]

Originally Posted by storm5510

I need to start looking at this by physical cores and not logical. So I changed the settings:

CoresPerTest=2
WorkerThreads=2

In the new snip below, the utilization is 100%, at the top-right. Below the graphs it is roughly half. Four threads are idling, but not cores. Temps are in the mid 60;s. The throughput from Prime95 is virtually unchanged.

See? Much better.

For what it's worth, your benchmarks show you'll get the most performance running a single worker. The difference is small but it will add up to an extra LL assignment per year if you leave Prime95 running all the time.

[storm5510]

Originally Posted by Mark Rose

See? Much better.

For what it's worth, your benchmarks show you'll get the most performance running a single worker. The difference is small but it will add up to an extra LL assignment per year if you leave Prime95 running all the time.

With this setup, Prime95 can do a DC in 36 hours. However, this is in the 40M range. I don't know what the server is passing out at this point.

I'm looking at the Exponent Status Distribution page on mersenne.org in another tab. I see a lot of available double-check's from 46M to 76M. PM1's are heavy from 82M through 146M.

I am going to make some changes here!

Update: I have changed my Prime95 configuration to do DC's. The server assigned three tests in the 45M range. Prime95 can complete each in 36 hours. The ms/iter is 2.9, give or take a very tiny bit. This is one worker and four cores per test. This seems to be the optimal spot.

[Madpoo]

Originally Posted by storm5510

With this setup, Prime95 can do a DC in 36 hours. However, this is in the 40M range. I don't know what the server is passing out at this point.

I'm looking at the Exponent Status Distribution page on mersenne.org in another tab. I see a lot of available double-check's from 46M to 76M. PM1's are heavy from 82M through 146M.

I am going to make some changes here!

Update: I have changed my Prime95 configuration to do DC's. The server assigned three tests in the 45M range. Prime95 can complete each in 36 hours. The ms/iter is 2.9, give or take a very tiny bit. This is one worker and four cores per test. This seems to be the optimal spot.

The double checks will be a great way to verify everything is working okay. The low end of double checks is currently working exponents in the 40-41M range.

You can look at the 4 (5 if we count zero) categories of assignments on this page and the rules on how long you have to finish, who qualifies to get what category of work, and a way to opt-in to getting "priority" work (smallest available exponents as long as you promise to finish them promptly).
https://www.mersenne.org/thresholds/