TL;DR version:
with this, and the previous compression level changes
I am seeing drastic speedups in package_write_rpm completion times:
webkitgtk goes from 78 seconds to 37 seconds
glibc-locale goes from 399 seconds to 58 seconds (!)
The long version:
rpm uses multithreading for two purposes:
- spawning compressors (which are nowadays themselves
multi-threaded, so the feature is not as useful as it once
was)
- parallel file classification
While the former behaves well on massively parallel CPUs
(it was written and verified here :), the latter was then added
by upstream and only benchmarked on their very old, slow laptop,
apparently:
https://github.com/rpm-software-management/rpm/commit/
41f0e214f2266f02d6185ba11f797716de8125d4
On anything more capable it starts showing pathologic behavior,
presumably from spawning massive amount of very short-lived threads,
and then having to synchronize them. For example classifying glibc-locale
takes
5m20s with 256 threads (default on my machine!)
1m49s with 64 threads
59s with 16 threads
48s with 8 threads
Even a more typical recipe like webkitgtk is affected:
47s with 256 threads
32s with 64 threads
27s with 16 or 8 threads
I have found that the optimal amount is actually four: this also
means that only four compressors are running at a time, but
as they're themselves using threads, and typical recipes are dominated
by just two or three large packages, this does not affect overall
completion time.
Signed-off-by: Alexander Kanavin <alex@linutronix.de>
Signed-off-by: Steve Sakoman <steve@sakoman.com>