Subject: Re: [netatalk-admins] Slow write performance on FreeBSD 2.2.5, 1.4b2
From: Bill Studenmund (skippy@macro.stanford.edu)
Date: Tue Feb 24 1998 - 14:57:27 EST
On Tue, 24 Feb 1998 wesley.craig@umich.edu wrote:
> > From: Bill Studenmund <skippy@macro.stanford.edu>
> > To: Gary Teter <garyt@bulldogbeach.com>
>
> > On Sat, 21 Feb 1998, Gary Teter wrote:
> > > Someone suggested changing the definition of ddp_recvspace in
> > > sys/netatalk/ddp_usrreq.c to a higher value, so I changed it to 50 * 600.
>
> > Doh! That kills my favorite suggestion!
>
> The normal value of ddp_recvspace, 10 * X, is enough for the eight
> packets you'll get back from an ATP transaction, plus a couple more for
> extra space. Each time this buffer is overflowed, there's a statistics
> block that gets updated, ddps_nosockspace, which your favorite BSD unix
> ought to return as output of netstat -s.
>
> Generally speaking, you should never see this...
>
> When your network performance is "really, really, slow", it's usually a
> bad network. AppleTalk tolerates a *very* low network error rate. In
> particular, write performance suffers because of the dual-ATP
> transactions used to "turn" the 8 packet response into a write. If
> you're getting, write, write, write, long ... pause ... write, write,
> etc, then it's almost certainly a bad network.
Wes,
Though you know the protocol better than I, I've had users clear up slow
write problems by upping ddp_recvspace to about 25 * X. I agree it
_shouldn't_ matter, but it has helped. I don't understand why (my
supposed reason doesn't jibe with your description above). I doubt going
to 50 would help much.
Take care,
Bill
This archive was generated by hypermail 2b28 : Sat Dec 18 1999 - 16:31:10 EST