[OpenAFS] Re: Optimizations

Andrew Deason adeason@sinenomine.net
Tue, 23 Apr 2013 10:12:08 -0500


On Tue, 23 Apr 2013 10:14:40 +0200
ProbaNet - Info <info@probanet.it> wrote:

> Everything works quite fine (except for some occasional slowdown when
> accessing some volume for the first time), but we would like to know
> if we can do something to improve it! :) Thank you very much!

If you're just talking about how fast a single client reads from the
fileserver, I don't think it's that likely that adjusting fileserver
options is going to do much (with the possible exception of -udpsize,
below). Client caching parameters may make more of a difference, but as
Lars said, posting the rates you are getting may help.

> P.S.: we don't understand the difference between "-udpsize" and
> "-sendsize" fileserver parameters: please can you explain it a bit? :)
> Thanks!

-sendsize alters the size of the buffer in the read/write loop in the
fileserver. That is, when we have something like this:

while (...) {
	rx_Read(buffer);
	write(buffer);
}

-sendsize is the size of the 'buffer'.

-udpsize just alters the size of the kernel's UDP packet buffer. That
is, how much data the kernel keeps in a buffer if we can't recvmsg()
fast enough. Adjusting -udpsize is probably more likely to result in a
noticeable difference than adjusting -sendsize. If you want to see if
the current -udpsize setting is not enough, you may find some guidance
here:
<http://openafs.org/pages/newsletter/newsletter-2013-03-volume004-issue05.html#_udpsize__socket_buffer_size_>.

-- 
Andrew Deason
adeason@sinenomine.net