Noticed my machines were running cooler than expected and checked the process manager and saw the long2's are only using 5 threads for long periods. They can drop down to 2 threads and increase thread count but never seem to ever make it to 6.
They are set to:
Is optimal 4 threads or less to maximize CPU usage?
I ran over deadlines leaving these long2's at 1 or 2 threads.
/ 
Reply
Quote
rebirtherVolunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Send message
Joined: 2 Jan 13
Posts: 7496
Credit: 44,029,375
RAC: 36,320
Noticed my machines were running cooler than expected and checked the process manager and saw the long2's are only using 5 threads for long periods. They can drop down to 2 threads and increase thread count but never seem to ever make it to 6.
They are set to:
<!-- srbase11 = Sierpinski / Riesel Base - long2 --> <app> <name>srbase11</name> <max_concurrent>1</max_concurrent> <fraction_done_exact/> </app> <app_version> <app_name>srbase11</app_name> <avg_ncpus>6</avg_ncpus> <cmdline>-t6</cmdline> </app_version>
Is optimal 4 threads or less to maximize CPU usage?
I ran over deadlines leaving these long2's at 1 or 2 threads.
Depends on how much CPU cache is used. On a ryzen 3950x there is a limit of 10 then it will be run slower.
/ Reply
Quote
Depends on how much CPU cache is used. On a ryzen 3950x there is a limit of 10 then it will be run slower.
3950x has 64mb lvl 3 so we should set the thread count to use no more than about 6mb per thread?
My older server Xeons only have 20MB (still better than my 2700X GPU host) so will have to limit the number of long2 WU's.
1 WU at 3 threads per CPU; currently they are running 2 WU at 6 threads per CPU package. (there's probably a slowdown of all other project WU's as they compete for cache w/ the long2's)
Are all the WU's needing this amount of cache?
If you have some data about max cache per WU type; I'll put it into a spreadsheet chart.
/ Reply
Quote
This is a management issue.
Resource share set to 0 on 16 core machine.
The project sends more longs than can be done in 1 day.
If I limit the long3 to 1 WU of 3 cores the project has sent 2, 3 or 4 WU and so 13 cores remain unused as the other connected projects are set to Resource share 0 and see a full work cache, so send no work. If I set one of those projects to resource share 1 then they dominate the work cache and SRBase sends no work.
I guess the only solution is to run SRBase within a 3 core VM's and put the rest of the workload in the main OS or other VM's.
BOINC's resource share system is terrible; designed when all computers were single core machines.
They need to add a work cache to each core, or group of cores, to give us more control over the workload.
/ Reply
Quote
rebirtherVolunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Send message
Joined: 2 Jan 13
Posts: 7496
Credit: 44,029,375
RAC: 36,320
This is a management issue.
Resource share set to 0 on 16 core machine.
The project sends more longs than can be done in 1 day.
If I limit the long3 to 1 WU of 3 cores the project has sent 2, 3 or 4 WU and so 13 cores remain unused as the other connected projects are set to Resource share 0 and see a full work cache, so send no work. If I set one of those projects to resource share 1 then they dominate the work cache and SRBase sends no work.
I guess the only solution is to run SRBase within a 3 core VM's and put the rest of the workload in the main OS or other VM's.
BOINC's resource share system is terrible; designed when all computers were single core machines.
They need to add a work cache to each core, or group of cores, to give us more control over the workload.
Thats true. The work cache is also based on the GFLOPs.
/ Reply
Quote