Hyperthreading

[Jud McCranie]

I have one computer that has hyperthreading. With the new version of Prime95 I notice that I can assign two workers. How much more throughput will I get by assigning two workers on a hyperthreaded CPU? (My guess 10-15%).

[starrynte]

On my Pentium 4, I get about 6% more throughput, but that varies depending on the programs I use that day.
Essentially, say 1 worker takes some amount of time. "Normally" it would take double that amount of time to run 2 workers. With hyperthreading though, 2 workers takes just slightly less than two times the amount of time. (Multithreading LL tests does work reasonably well on hyperthreading though)

[petrw1]

Originally Posted by starrynte

On my Pentium 4, I get about 6% more throughput, but that varies depending on the programs I use that day.
Essentially, say 1 worker takes some amount of time. "Normally" it would take double that amount of time to run 2 workers. With hyperthreading though, 2 workers takes just slightly less than two times the amount of time. (Multithreading LL tests does work reasonably well on hyperthreading though)

Not me. I have two single-core hyperthreading PIV's; a 2.4 Ghz and a 3.4 Ghz. Both actually run about 3-5% SLOWER when I choose 2 CPUs.

I have NOT tried v25.9, though.

[starrynte]

Originally Posted by petrw1

Both actually run about 3-5% SLOWER when I choose 2 CPUs.

This occasionally happens with mine, but 90% of the time that's because some program I'm running is "stealing" CPU cycles. So assuming average usage, you'd get anywhere from 10% increase to 5% decrease in performance, depending on the programs being used...

[Jud McCranie]

Thanks, I have enough workers running that I don't think using the hyperthreading is really worth it for me.

[Maybeso]

On my P4 I found that running one LL thread and one TF thread gave me at least +15%. The task manager would show 100% CPU. In the evenings when I was usually home, I had the TF thread stop and task manager showed 50% each thread, as expected.

Now on version 25, it shows 100% with only one worker doing just LL. Is the LL using both threads better via the "helper"? If I can specify different work types for each worker I will try using 2. Otherwise I don't see how to do it.

[starrynte]

<assuming you are doing DC/LL tests>
If you set 1 worker, multithreading 1, I believe there will be 1 worker using both logical cores (for a total of 1 thread working on the assignment, there's no helper thread!). This, naturally, uses all the CPU. (Though it may appear as only 50%, I've read somewhere that this was because of hyperthreading, but I'm not completely sure)
If you set 1 worker, multithreading 2, there will be 1 worker along with a helper thread (for a total of 2 threads working on the assignment). This does NOT use 100% CPU (at least not on mine), only something between 80 and 90. (100% CPU usage may occur though, if you have a lot of other programs running)
If you set 2 workers, multithreading 1, there will be 2 workers (for a total of 1 thread working on each worker's assignment). This uses 100% CPU.
If you set 2 workers, multithreading 2, Prime95 will warn you as this means each worker is using 2 threads, for a total of 4 threads. Yet there are only 2 CPU's for 4 threads to run on, meaning DECREASED performance.

Is the LL using both threads better via the "helper"?

An LL with a helper = 1 worker, multithreading 2. In the long run, performance will be decreased (slightly), but if all you want to do is finish the assignment, a helper thread will speed up the LL. (Though be warned that if the helper thread is affected in performance, the main thread is also.)

If I can specify different work types for each worker

You can set a different work type for each worker that you have.

I have a Intel I7, which is a quad core with hyperthreading. So right now I'm running 4 workers at 2 multithreading each, as my cpu has 8 logical cores with hyperthreading on and my cpu usage is at 50%. Shouldn't it be at least around 80%-100%

[starrynte]

oh and also regarding apps "stealing" cpu cycles, for those who use Digsby (an IM client) there is apparently a "research module" that runs distributed computing projects after the computer has been idle for 5 minutes. (see http://blog.digsby.com/archives/68 for more details) To disable, go to Help -> Support Digsby and disable "Help Digsby Conduct Research" (enabled by default thus far)

[xorbe]

Originally Posted by Unregistered

I have a Intel I7, which is a quad core with hyperthreading. So right now I'm running 4 workers at 2 multithreading each, as my cpu has 8 logical cores with hyperthreading on and my cpu usage is at 50%. Shouldn't it be at least around 80%-100%

I think TF only uses 1 thread. If you do LL, then it will use any number of threads I think. Not sure about P-1, as I only did one of those so far.

[cheesehead]

Originally Posted by xorbe

I think TF only uses 1 thread. If you do LL, then it will use any number of threads I think. Not sure about P-1, as I only did one of those so far.

Seems to me that beneficial multithreading ought to be a matter of whether or not FFTs are involved. That is, the same benefit for LL would apply to P-1 and ECM, so thread-splitting could be deferred to the FFT code common to all three. I've looked briefly for it in the source code, but haven't found where thread-splitting is done.