View Single Post
Old 2020-03-07, 16:42   #8
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

11×383 Posts
Default

Quote:
Originally Posted by kuratkull View Post
(running all this on a Skylake i7)

Using zero-padded FMA3 FFT length 384K for both:
4 threads: 39547695*2^3664022-1 is not prime. LLR Res64: DA225779C3F421FD Time : 1409.882 sec.
3 threads: 39547695*2^3664034-1 is not prime. LLR Res64: 7F70D51694B8E6AF Time : 1645.871 sec.
384 / 4 = 96
384 / 3 = 128
I would have expected it to work better with 3 threads. Will have to look into optimal settings/recommendations for the user.
4 * 1409 = 5636 thread-sec
3 * 1645 = 4935 thread-sec

I consider 15% more thread-sec 'inefficient'. I certainly didn't make clear what I meant the first time, sorry! If we did a similar comparison for 2- vs 3- threaded, I imagine the 2-threaded instance would be ~4800 thread-sec, within 3-4% of the 3-threaded instance. That's a tradeoff I'm willing to make for the admin simplicity of fewer clients running, but 15% is not.
VBCurtis is offline   Reply With Quote