Measuring overheads of some threading API¶
You can find the code for benchmarking and plotting at https://github.com/tkf/ThreadsAPIBenchmarks.jl.
using DisplayAs
using ThreadsAPIBenchmarksAnalyzers: TaskOverheads, NotifyOverheads
Overhead of @spawn and @sync¶
Measuring the overhead of @spawn and @sync is not entirely trivial.
@btime wait(Threads.@spawn nothing) does not reflect the actual overhead
since @spawn may use the current thread to run the task. That is to say,
the following check likely to pass even if nthreads() > 1:
let n = 0
for _ in 1:1000
n += fetch(Threads.@spawn Threads.threadid()) == Threads.threadid()
end
@assert n > 0
end
ThreadsAPIBenchmarks.jl contains a benchmark for measuring the overheads of
@sync and @spawn. That is to say, it measures @spawn overhead t1 - t0
and wait (sync) overhead t2 - t1 of
t0 = time_ns()
wait(@spawn t1 = time_ns())
t2 = time_ns()
while making sure @spawned task is executed in a different OS threads.
The distribution of the above measurement is shown below for different number
N = nthreads() of threads:
begin
task_overheads = TaskOverheads.analyze()
task_overheads.hists
end |> DisplayAs.PNG
Here is a more direct plot of the raw data to show that the multimodality of the distribution comes from the temporal proximity of the measurements (some state transitions in Julia runtime, OS, or hardware?):
begin
# task_overheads = TaskOverheads.analyze()
task_overheads.seq
end |> DisplayAs.PNG
Overhead of notify (schedule)¶
Following distribution describes the overheads of wait and notify on
Threads.Condition. It is essentially the overhead of re-schedule.
The first two rows (max and min) shows the distribution of the time took
for notify to end the wait of the peer threads. The first row max shows
the slowest among all threads and second row min is for the fastest. The
last row notify is @time notify(...).
begin
notify_overheads = NotifyOverheads.analyze()
notify_overheads.hists
end
Compared to @sync and @spawn, there are not much temporal irregularities
especially for N > 2:
begin
# notify_overheads = NotifyOverheads.analyze()
notify_overheads.seq
end |> DisplayAs.PNG
This notebook was generated using Literate.jl.