We currently use the semaphore of our sleep strategy very greedy.
This is due to skipping the semaphore's V() operation if we are sure
that it does not kill our progress guaranty.
When scheduling new work from within the runtime we skip the wakeup if we observe nobody sleeping. This is fine in terms of progress and limits the amount of atomic operations on global state to a minimum.
Using a threshold of 0 (observe nobody sleeping) however introduces a race between inserting new work and going to sleep which harm the latency when the worker goes to sleep and is not notified about the new work.
This race is common and can be observed in the pulse micro benchmark. A emper with a threshold of 0 shows high latency compared to using an io-based sleep strategy or when increasing the threshold.
$ build-release/eval/pulse | python -c "<calc mean>"
Starting pulse evaluation with pulse=1, iterations=30 and utilization=80
mean: 1721970116.425
$ build-increased-sem-threshold/eval/pulse | python -c "<calc mean"
Starting pulse evaluation with pulse=1, iterations=30 and utilization=80
mean: 1000023942.15
$ build-pipe-release/eval/pulse | python -c "<calc mean>
Starting pulse evaluation with pulse=1, iterations=30 and utilization=80
mean: 1000030557.0861111
$ build-pipe-no-completer/eval/pulse | python -c "<calc mean>"
Starting pulse evaluation with pulse=1, iterations=30 and utilization=80
mean: 1000021514.1805556I could not measure any significant overhead due to the more atomics on my 16 core machine using the fs-eval on a SSD.
$ ./eval.py -r 50 -i emper-vanilla emper-inc-sem-threshold emper-pipe emper-pipe-no-completer
...
$ ./summarize.py results/1599f44-dirty-pasture/<date>/ -f '{target}-{median} (+- {std})'
duration_time:u:
emper-vanilla-0.202106189 (+- 0.016075981812486713)
emper-inc-sem-threshold-0.2049344115 (+- 0.015348506939891596)
emper-pipe-0.21689131 (+- 0.015198438371285145)
emper-pipe-no-completer-0.1372724185 (+- 0.005865720218833998)