Not nice enough

[phma]

 4427 phma      30  10 1091452 424224   4984 S 430.9   1.3  93463:58 mprime
28163 phma      20   0 3588756   1.8g  80440 R 412.6   5.7 206:19.02 perfecttin-gui

I'm running Linux 5.3.0 on a 6-core 12-thread AMD box. Mprime is running six worker threads, all working on the same Mersenne number. PerfectTIN is a program I wrote for work. It has twelve worker threads, all building the same TIN out of the same point cloud.

When I'm running PerfectTIN, I'd like it to take nearly all of the CPU. I've seen it take 80 to 98% of the CPU without mprime running. (It has a lock contention bottleneck, so when it's adding triangles as fast as possible, the CPU usage goes down a bit.) How can I make mprime take less CPU when other high-CPU programs are running?

[M344587487]

nice works with values from -20 to 19, you could try running mprime with niceness of 19 and PerfectTIN with niceness -20.

[retina]

Originally Posted by phma

How can I make mprime take less CPU when other high-CPU programs are running?

Add PauseWhileRunning to your config.

The problem you see isn't with nice values, but the OS scheduler. Windows will do a better job of scheduling within the priorities given, but I wouldn't recommend changing to that.

[Ensigm]

Originally Posted by phma

When I'm running PerfectTIN, I'd like it to take nearly all of the CPU.

PauseWhileRunning is probably what you will need. Alternatively, you can set MaxLoad=A and MinLoad=B so that mprime will suspend a worker whenever the system load gets above A and resume a worker when it goes below B. You can start out with something like MaxLoad=8 and MinLoad=6 and try different combinations to see what works best. Nice value is not of much significance here.

[Ensigm]

For more detailed information, check undoc.txt.

[Mark Rose]

Originally Posted by M344587487

nice works with values from -20 to 19, you could try running mprime with niceness of 19 and PerfectTIN with niceness -20.

That's a great way to lock up your machine. Anything that saturates the CPU should never be given a priority higher than 0.

[Mark Rose]

Originally Posted by phma

 4427 phma      30  10 1091452 424224   4984 S 430.9   1.3  93463:58 mprime
28163 phma      20   0 3588756   1.8g  80440 R 412.6   5.7 206:19.02 perfecttin-gui

I'm running Linux 5.3.0 on a 6-core 12-thread AMD box. Mprime is running six worker threads, all working on the same Mersenne number. PerfectTIN is a program I wrote for work. It has twelve worker threads, all building the same TIN out of the same point cloud.

When I'm running PerfectTIN, I'd like it to take nearly all of the CPU. I've seen it take 80 to 98% of the CPU without mprime running. (It has a lock contention bottleneck, so when it's adding triangles as fast as possible, the CPU usage goes down a bit.) How can I make mprime take less CPU when other high-CPU programs are running?

Unfortunately mprime doesn't set the scheduling priority to SCHED_IDLE on Linux. You can do that manually by running `chrt -i -p 0 ###` where ### is mprime's process id.

There was some recent work by Facebook on fixing sched-idle in the CFS scheduler, but I don't know if it's made it into mainline kernels. You might have better performance using a different scheduler like MuQSS or BFS, but that's really only something you should get into if you're comfortable changing kernels and so on.

[M344587487]

Originally Posted by Mark Rose

That's a great way to lock up your machine. Anything that saturates the CPU should never be given a priority higher than 0.

That's interesting, can't remember any setting having too much of an impact but I guess I didn't play around with it enough.

[Xyzzy]

With the weird way that a negative nice value is "positive" it is real easy to accidentally give a process the wrong value.

As for the computer becoming unresponsive? Yep!

[Mark Rose]

Originally Posted by Xyzzy

With the weird way that a negative nice value is "positive" it is real easy to accidentally give a process the wrong value.

As for the computer becoming unresponsive? Yep!

Yeah, remember negative numbers are less nice than normal, not more nice. The semantic is backwards.

[phma]

Originally Posted by Mark Rose

Unfortunately mprime doesn't set the scheduling priority to SCHED_IDLE on Linux. You can do that manually by running `chrt -i -p 0 ###` where ### is mprime's process id.

I tried that; it doesn't seem to make a difference.

Originally Posted by retina

Add PauseWhileRunning to your config.

That would keep mprime stopped as long as PerfectTIN was running, even after it had finished processing.

The problem you see isn't with nice values, but the OS scheduler. Windows will do a better job of scheduling within the priorities given, but I wouldn't recommend changing to that.

I'm running Linux. What Windows would do is irrelevant. So is BSD (and the top output would look different).