using multiple threads on an LL assignment

tha · 2016-01-31, 13:42

I have a Skylake i7, so four real cores and eight logical cores for hyperthreading.

If I set number of worker threads to 4 the throughput is four LL tests, each taking 14.8 milliseconds/iteration.

If I set number of worker threads to 2 and the number of CPUs to use in LL test (multithreading) to 2 the throughput is two LL tests, each taking 7.7 milliseconds/iteration.

I have set worker 1 to use core 1 out of 1..8 and worker 2 to use core 3 out of 1..8, but Mprime then seams to override this and assigns the first worker to logical CPU#1 to be helped by logical CPU#5 and assigns the second worker to logical CPU#2 and helped by logical CPU#6

Next I have set the number of workers set to 2 and the number of logical cores per worker to be used to 4. All using smart assignment. Now the timings are down to 4.6 milliseconds/iteration.

(On a side note: why does part of the program number the cores form 0..7 and other parts 1..8)

tha · 2016-01-31, 19:36

Quote:

Originally Posted by tha

Next I have set the number of workers set to 2 and the number of logical cores per worker to be used to 4. All using smart assignment. Now the timings are down to 4.6 milliseconds/iteration.

And when setting all other 7 threads to help the first one the timings for the same exponent (2240K FFT) is down to 2.3 millseconds. It really speeds up things in this setting. I guess because of the cache memory having more hits. All cores are now +/- 96% busy.

ATH · 2016-01-31, 19:48

You should disable hyperthreading in the BIOS if possible. There is no value in hyperthreading for Prime95, it is mostly for "normal" lightweight stuff like office programs and browsing that can easily run 2 threads on 1 core.

If you run a benchmark with these 2 lines in prime.txt:
BenchMultipleWorkers=1
BenchMultithreads=1

It will test like 4 workers 1 core on each, 2 workers 2 cores on each and 1 worker with 4 cores and give you a total iteration/sec for each one, so you can see what gives the largest throughput.

Madpoo · 2016-02-01, 05:20

Quote:

Originally Posted by ATH

You should disable hyperthreading in the BIOS if possible. There is no value in hyperthreading for Prime95, it is mostly for "normal" lightweight stuff like office programs and browsing that can easily run 2 threads on 1 core.

If you run a benchmark with these 2 lines in prime.txt:
BenchMultipleWorkers=1
BenchMultithreads=1

It will test like 4 workers 1 core on each, 2 workers 2 cores on each and 1 worker with 4 cores and give you a total iteration/sec for each one, so you can see what gives the largest throughput.

And also this to disable even bothering throwing HT cores into the mix:
BenchHyperthreads=0

rudi_m · 2016-02-02, 13:49

FYI I've noticed that I get the best throughput compromise when using 3 threads for LL plus 1 thread for factoring. The LL threads have more bad influence against each other than factoring threads.

BTW I've noticed that the CPU temperature may differ dependently on what kind of work runs on which specific CPU core. For example I have such a worktodo.txt, which gives me the best CPU temperature.

Code:

[Worker #1]
Factor=...

[Worker #2]
DoubleCheck=...

[Worker #3]
Factor=...

[Worker #4]
DoubleCheck=...

Any other permutation of the jobs gives me much higher temperature and causes the CPU to throttle (cooling is bad on this machine).

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Multiple threads per assignment?	f0rteOC	Hardware	3	2016-02-29 06:08
Multiple Core for 1 assignment	saeres	Information & Answers	10	2016-02-04 01:52
Workers, Threads, Helper Threads, Cores, Affinity.	lorgix	Information & Answers	12	2011-01-13 22:31
pearl question: multiple threads of ECM	Andi47	Programming	2	2010-09-25 06:34
Multiple threads efficiency...	Xyzzy	Linux	1	2004-12-31 21:03

2016-01-31, 13:42	#1
tha Dec 2002 32E₁₆ Posts	using multiple threads on an LL assignment I have a Skylake i7, so four real cores and eight logical cores for hyperthreading. If I set number of worker threads to 4 the throughput is four LL tests, each taking 14.8 milliseconds/iteration. If I set number of worker threads to 2 and the number of CPUs to use in LL test (multithreading) to 2 the throughput is two LL tests, each taking 7.7 milliseconds/iteration. I have set worker 1 to use core 1 out of 1..8 and worker 2 to use core 3 out of 1..8, but Mprime then seams to override this and assigns the first worker to logical CPU#1 to be helped by logical CPU#5 and assigns the second worker to logical CPU#2 and helped by logical CPU#6 Next I have set the number of workers set to 2 and the number of logical cores per worker to be used to 4. All using smart assignment. Now the timings are down to 4.6 milliseconds/iteration. (On a side note: why does part of the program number the cores form 0..7 and other parts 1..8) Last fiddled with by tha on 2016-01-31 at 14:01

2016-01-31, 19:48	#3
ATH Einyen Dec 2003 Denmark 3⁵×13 Posts	You should disable hyperthreading in the BIOS if possible. There is no value in hyperthreading for Prime95, it is mostly for "normal" lightweight stuff like office programs and browsing that can easily run 2 threads on 1 core. If you run a benchmark with these 2 lines in prime.txt: BenchMultipleWorkers=1 BenchMultithreads=1 It will test like 4 workers 1 core on each, 2 workers 2 cores on each and 1 worker with 4 cores and give you a total iteration/sec for each one, so you can see what gives the largest throughput. Last fiddled with by ATH on 2016-01-31 at 19:53

2016-02-02, 13:49	#5
rudi_m Jul 2005 2·7·13 Posts	FYI I've noticed that I get the best throughput compromise when using 3 threads for LL plus 1 thread for factoring. The LL threads have more bad influence against each other than factoring threads. BTW I've noticed that the CPU temperature may differ dependently on what kind of work runs on which specific CPU core. For example I have such a worktodo.txt, which gives me the best CPU temperature. Code: [Worker #1] Factor=... [Worker #2] DoubleCheck=... [Worker #3] Factor=... [Worker #4] DoubleCheck=... Any other permutation of the jobs gives me much higher temperature and causes the CPU to throttle (cooling is bad on this machine).