In order to know when to switch to "no new tasks", I checked the min/avg/max run times of GFN21 on the different host types which I am using in this challenge, using a spreadsheet. As a byproduct I calculated PPD from the average run times and average credits of all currently validated tasks. (All hosts run Linux. The two E5 hosts don't run GPU tasks.)
dual E5-2690 v4: 700,000 PPD
dual E5-2696 v4: 980,000 PPD
dual GTX 1080Ti: 810,000 PPD
In comparison, the dual E5-2690 v4 gets only 370,000 PPD in
GCW-Sieve and 165,000 PPD in
PPS-LLR for example.
I suppose PPD and efficiency of
cpuGFN21 are dramatically lower on CPUs which have less than ~25 MB L3 cache. Naturally I have such CPUs too, but I kept them at World Community Grid.
Edit:
OK, maybe not dramatically, but distinctly. Taking data from
post #608 and assuming 149,000 credits per task on hosts which I did not test myself:
dual E5-2690 v4 @ 2.9 GHz, 2x quadchannel DDR4-2400, 2 tasks x 14 threads/task:
350,000 PPD per socket
25,000 PPD per physical core
8,600 PPD per physical core and per GHz
dual E5-2696 v4 @ 2.6 GHz, 2x quadchannel DDR4-2400, 4 tasks x 11 threads/task:
490,000 PPD per socket
22,000 PPD/core
8,600 PPD/core/GHz
i5-8600K @ 4.3 GHz, dualchannel DDR4-????, 1 task x 4 threads*:
121,000 PPD per socket
20,000 PPD/core counting 6 cores, or 30,000 PPD/core counting 4 cores
4,700 PPD/core/GHz counting 6 cores, or 7,000 PPD/core/GHz counting 4 cores
Ryzen 2700X @ 4.1 GHz, dualchannel DDR4-3466, 1 task x 8 threads:
123,000 PPD per socket
15,400 PPD/core
3,800 PPD/core/GHz
*) On Intel CPUs with 4 cores or more, but small cache and dual channel RAM,
cpuGFN21 reportedly scales only up to 4 threads per task, and throughput remains the same or suffers if more threads than that are used, even if there are more cores to spare.