On 2019-06-24 22:38, Linus Torvalds wrote:
On Mon, Jun 24, 2019 at 10:42 PM Roman Penyaev <rpenyaev@xxxxxxx>
wrote:
So harvesting events from userspace gives 15% gain. Though bench_http
is not ideal benchmark, but at least it is the part of libevent and
was
easy to modify.
Worth to mention that uepoll is very sensible to CPU, e.g. the gain
above
is observed on desktop "Intel(R) Core(TM) i7-6820HQ CPU @ 2.70GHz",
but on
"Intel(R) Xeon(R) Silver 4110 CPU @ 2.10GHz" measurements are almost
the
same for both runs.
Hmm. 15% may be big in a big picture thing, but when it comes to what
is pretty much a micro-benchmark, I'm not sure how meaningful it is.
And the CPU sensitivity thing worries me. Did you check _why_ it
doesn't seem to make any difference on the Xeon 4110? Is it just
because at that point the machine has enough cores that you might as
well just sit in epoll() in the kernel and uepoll doesn't give you
much? Or is there something else going on?
This http tool is a singlethreaded test, i.e. client and server
work as a standalone processes and each has a single event thread
for everything.
According to what I saw there, is that events come slowly (or event
loop acts faster?), so when time has come to harvest events there
is nothing, we take a slow path and go to kernel in order to sleep.
That does not explain the main "why", unfortunately.
I would like to retest that adding more clients to the server, thus
server is more likely to observe events in a ring, avoiding sleep.
--
Roman
