As folks may have noticed, I've been re-working my old 2015 dispatch
patches that eliminate the network input-side queues in Ganesha.
Matt had wanted fully async non-blocking I-O. I've been poking at it
for a week, and now am sure that's the wrong way to go.
It might still be good for FSALs. Remains to be seen. DanG and
Soumya are looking at that now.
The devil in userland network I-O is system calls. Each epoll_wait
is a system call. Each read or write is a system call. Each thread
switch is a system call.
My code in Ganesha v2.5 (NTIRPC v1.5) gets the network output down to
one system call per request on a very hot thread. Cannot do better,
as trying harder would just push the data into kernel buffers,
possibly slowing our own output (for various reasons).
Trying to re-work that for async non-blocking calls instead means
many more system calls. Instead of one clean writev with the TCP
fragment header and all ready buffers in one single call, we'd at
minimum have a call, an epoll_wait, spawn another work thread, then
another call and/or release the buffer, rinse and repeat.
For a long buffer chain (the times we want more performance), we'd
have much less performance -- roughly 2 + (3 * number of buffers)
additional system calls. For common short response chains, still
have the extra overhead of the epoll system call, doubling calls.
Also, using writev minimizes buffer copies. Eliminating data
copying will usually give far better performance.
The only thing async output is saving is waiting threads. But I've
already got the output threads down to the minimum (per interface).
No gain here!
On the input side, the truly optimum reduction in system calls would
be one read to get the TCP fragment header and up to 1500 bytes of
data, followed (only when needed) by another read to get the entire
rest of long fragments in one fell swoop.
With async input I've tried level triggered, and am getting spurious
epoll read data signals. Googling shows that's been a problem since
at least 2014, but possible to program around.
Still, this could be better, had it not been terrible for output-side.
Changing to edge triggered means that every good read would be
followed by another read to make sure that we've gotten all the data.
That is, common small reads turn into two (2) reads. Doubling our
system calls in the common case is not the way to go....
In conclusion, with epoll we know when input data is available, so
input threads aren't sitting around waiting anyway, and trying to
minimize threads results in more system calls and poorer performance.
NTIRPC already defaults to 200 worker threads. If we need more, we
should allocate more. Memory should not be an issue.
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
Nfs-ganesha-devel mailing list