View Single Post
Old 2021-09-23, 17:55   #439
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

3×11×179 Posts
Default prime95 idles lots of cores on a worker during P-1 GCD

See attachment. Same will apply to dual-manycore-Xeons and some other configurations.
P-1 during GCD idles all but 1 core of a worker. On a Xeon Phi this may be 64/n -1 or 68/n -1 cores, where n is number of workers. Depending on exponent the duration may be considerable.
Gpuowl handled this situation by running parallel threads, speculatively executing the next P-1 stage or the following PRP while the GCD ran in a separate thread. The factor finding probabilities with normal bounds are such that ~98% of the time the same exponent is involved, it pays off. And if it is a succession of P-1 on different exponents, or following is other type work on different exponents, whatever payoff there is occurs 100% of the time. Perhaps this approach would be productive on hyperthreaded CPUs in prime95 / mprime also. (And Mlucas?)
In the M880M case I was running, it took an hour for a GCD.

Similarly, during preallocation of disk space for the proof residues of an M500M PRP, it took a few minutes while nearly all cores of the worker were idle.
Attached Thumbnails
Click image for larger version

Name:	most worker cores idle during P-1 GCD.png
Views:	43
Size:	54.4 KB
ID:	25743  
kriesel is offline   Reply With Quote