First of all this is a gaming PC I built last year but quickly ran out of time to use it for its intended purpose. Now it crunches 24/7 with a weekly reboot on Saturday.

I've been looking to optimize performance and have noticed all tasks sit at 0% for exactly 45 seconds then finish normally a couple minutes later. Is this normal behavior?

Currently I'm playing around with how many cores to allocate per task and seem to be experiencing diminishing returns with actual task start time. For example... if all tasks finish around the same time the CPU sits idle until the next ones actually start. While monitoring RAM the bandwidth ranges from 1.5-4GBPS.while running and sub 1GBPS until they actually start. Any suggestions?

I've been looking to optimize performance and have noticed all tasks sit at 0% for exactly 45 seconds then finish normally a couple minutes later. Is this normal behavior?

Yes.

Currently I'm playing around with how many cores to allocate per task and seem to be experiencing diminishing returns with actual task start time. For example... if all tasks finish around the same time the CPU sits idle until the next ones actually start. While monitoring RAM the bandwidth ranges from 1.5-4GBPS.while running and sub 1GBPS until they actually start. Any suggestions?

Use one thread per task. This eliminates the times when the CPU sits nearly idle during the single thread start up phase of each task.

According to my calculation for my Ryzen system, the multi-threaded application (v1.86) using just two threads per task would generate a RAC of about 14k, the single-threaded (v1.86) about 22k, however I think there seems to be some CreditNew lottery involved considering the lower credits for v1.87 (RAC ~20k), so YMMV, but I expect one thread per task to be most efficient in any case. And my CPU also seems to be using about 4 Watts less when running 16x 1 thread instead of 8x 2 thread, no idea why, but it does.

---

I've been looking to optimize performance and have noticed all tasks sit at 0% for exactly 45 seconds then finish normally a couple minutes later. Is this normal behavior?

Currently I'm playing around with how many cores to allocate per task and seem to be experiencing diminishing returns with actual task start time. For example... if all tasks finish around the same time the CPU sits idle until the next ones actually start. While monitoring RAM the bandwidth ranges from 1.5-4GBPS.while running and sub 1GBPS until they actually start. Any suggestions?

i don't think the cores are idle just waiting for Boinc to recognize the lowest percentage of crunching that MW has programmed into the tasks, somewhere around 1 to 2%, this is also where the task checks to see if it's been run previously and has a checkpoint to start from etc etc etc. You can look at the logs if you really want to, all Boinc tasks from every Project do similar things, by looking in the stderr file in the boinc/slots directories.

I agree with Link 1 core per task seems to work for me too.

i don't think the cores are idle just waiting for Boinc to recognize the lowest percentage of crunching that MW has programmed into the tasks, somewhere around 1 to 2%

Yes, the cores are idle, the initial phase is single core, so if you run Milkyway with the default 16 threads per task, 1 will be used and 15 idle. There's also no progress during that phase reported by the application, it stays at 0.000% until it's done with whatever it's doing there on that single core.

---

i don't think the cores are idle just waiting for Boinc to recognize the lowest percentage of crunching that MW has programmed into the tasks, somewhere around 1 to 2%

Yes, the cores are idle, the initial phase is single core, so if you run Milkyway with the default 16 threads per task, 1 will be used and 15 idle. There's also no progress during that phase reported by the application, it stays at 0.000% until it's done with whatever it's doing there on that single core.

I don't see any of that because I run them all single core.

Yes, that eliminates the issue completely.

---

Great catch btw.

I switched from 16 cores to 2 cores per WU. Odd that the delay went from ~30 sec to ~60 sec so the other cores were doing something "during idle" when 16 were process each WU.

With 32 cores, I have 16 running now.

Maybe I'll try one core per WU later but it seems to be much more efficient per task manager.

Odd that the delay went from ~30 sec to ~60 sec

That's expected, now the start up phase has to share a physical CPU core with one thread from another task, before it had the whole core for itself.

---

I did switch to from 2 to 1 core per work unit. Thanks again.