[2145] in SIPB-AFS-requests
Re: afsd priorities?
daemon@ATHENA.MIT.EDU (Richard Basch)
Mon Sep 18 14:15:37 1995
Date: Mon, 18 Sep 1995 14:10:00 -0400
To: Derek Atkins <warlord@MIT.EDU>
Cc: John Hawkinson <jhawk@MIT.EDU>, sipb-afsreq@MIT.EDU
In-Reply-To: Derek Atkins's message of Sat, 16 Sep 1995 18:32:00 EDT,
<199509162232.SAA29406@toxicwaste.media.mit.edu>
From: "Richard Basch" <basch@MIT.EDU>
As Derek said, it really depends on the kernel scheduling mechanism.
Theoretically, you might get some performance improvement with the
readahead feature of AFS, since that will run more frequently. The
normal I/O should only be marginally improved since all that will happen
when one process blocks waiting for data from AFS and the afsd threads wakeup.
Raising the priority simply should marginally improve the wakeup a
little more often, but it is unlikely to make a substantial improvement.
I would imagine that if you are reading a lot of sequential data, the
raised priority of the threads doing the readahead is what is helping.
Anyway, just be aware that the raised priority for the readaheads can
work against you too -- making readaheads more likely than other I/O
from other sources...
This one is a hard call... for the dialups I can well imagine that the
priority being raised will help performance as the bulk of the traffic
is filesystem traffic, but for something like a server running multiple
services such as IRC and something else that performs a lot of AFS I/O,
you may find IRC and other similar services degraded slightly...
Probably by default in the environment, you won't notice much difference
or perhaps a slight improvement in performance because the bulk of the
work done by Athena users is file i/o (excluding local computation).
-Richard
