Ubuntu Generic vs. Low-Latency Linux Kernel Benchmarks For HPC & Desktop

[dsmythies]

I am curious about the Stress-NG 0.16.04: Socket Activity test results, which seems to be somewhat of an outlier data point.
lowlatency: 13,469.98 Bogo Op/s
generic: 29,718.94 Bogo Op/s

I increased the test runs from 3 to 10 and from 30 to 60 seconds per run and got:
250Hz kernel (generic): 6608.74 Bogo Op/s, Deviation 0.56%
1000Hz kernel (lowlatency): 6591.27 BOgo Op/s, Deviation 0.39%

My kernel was mainline 6.8-rc1 using Ubuntu kernel configurations. Processor: Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz
There was no power or temperature or number of active cores throttling during the test, although it was close to both the power and temperature limits set for my system.

[indepe]

I am curious about the Stress-NG 0.16.04: Socket Activity test results, which seems to be somewhat of an outlier data point.
lowlatency: 13,469.98 Bogo Op/s
generic: 29,718.94 Bogo Op/s

I increased the test runs from 3 to 10 and from 30 to 60 seconds per run and got:
250Hz kernel (generic): 6608.74 Bogo Op/s, Deviation 0.56%
1000Hz kernel (lowlatency): 6591.27 BOgo Op/s, Deviation 0.39%

My kernel was mainline 6.8-rc1 using Ubuntu kernel configurations. Processor: Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz
There was no power or temperature or number of active cores throttling during the test, although it was close to both the power and temperature limits set for my system.

If that's a throughput test which is latency-independent, the difference of 0.27% sounds more plausible.

With latency-independent I mean this: the test could be more latency-only by having a ping-pong test of 1-byte messages between 2 threads (on the same core), or a mixed test by using a ping-pong of larger messages (for example 10 kByte each).

In any case, I'd be curious about the results of a ping-pong Unix-socket test with both small and large messages. Will they be as expected?

EDIT: Actually I'm not quite sure what would the best way to test the "success" of latency-only.

[Volta]

However I'd expect the responsiveness to be at least slightly better, even if it might *perhaps* not be perceptible. But why would it be worse? What is a possible reason this could cause 1-2 second pauses in the UI ?

I'm guessing it may be related to GPU/rendering. Maybe there's some locking messed up in desktop environment/display server or maybe it's related to I/O. There's many possibilities that should be traced. Let's say CPU did its part, but can't continue, because GPU wants something else from it. Or vice versa. With longer slices everything finishes its job on time perhaps, so its ready to render. With lower latency they may be divided too much - want to print window content, but it's being preempted by some other process and so on. Just guessing. Probably could be solved with smart process grouping and appropriate priorities for user important tasks.

[indepe]

I'm guessing it may be related to GPU/rendering. Maybe there's some locking messed up in desktop environment/display server or maybe it's related to I/O. There's many possibilities that should be traced. Let's say CPU did its part, but can't continue, because GPU wants something else from it. Or vice versa. With longer slices everything finishes its job on time perhaps, so its ready to render. With lower latency they may be divided too much - want to print window content, but it's being preempted by some other process and so on. Just guessing. Probably could be solved with smart process grouping and appropriate priorities for user important tasks.

Although you are not naming anything specific, it sounds like you have something in mind that might explain missing a time slice or two per frame, or so. But a pause of 1-2 seconds? I'd think something specific must be going wrong that can be fixed.

[JosiahBradley]

I feel old and way out the loop again. I thought we got rid of the tick system over a decade ago. It's been awhile since I did a kernel config fully but I remember being able to choose between 1000, 500, 250, 100, or tickless. Always used tickless on my desktop and 100 on servers. I kinda want to find time to do another kernel build after all these years working instead of testing.