About a week ago, I wrote asking about the above subject - since I think
I got some good advice, and a couple of metoo's, here is a short summary.
Basically, the issue was how to configure ODS to get the most performace
improvement out of using two disks instead of one for spool.
Everybody agreed that using striping with a small interlace was the
absolute worst thing to do - just about every article access would have
to request I/O from both disks, with little or no benefit (the only case
were it makes sense to have an interlace smaller than typical file sizes
is with single processes seqentially accessing very large files - i.e.
the exact opposite of News).
As for striping with a large interlace (i.e. cylinder size) vs
concatenation, opinions varied a bit. Several respondents agreed with my
suspicion that it wouldn't really matter much one way or the other, but
Jerry Aguirre pointed out a couple of possible advantages with striping:
Since the FFS tries to keep directories (=newsgroups) "together", you
would tend to get all of a given newsgroup on one of the disks with
concatenation, whereas it would be more likely to be split between them
with striping - this means that with concatenation,
- you *could* end up with unbalanced load, e.g. if one of the disks
happened to get a higher proportion of high traffic/readership groups
- expiration (i.e. INN's fastrm) would typically beat at one disk at a
time, rather than splitting the load between them.
Kean Stump had actually tried both striping and concatenation for the
News spool, and gotten much better performance with concatenation. He
had used a small (4k) interlace though, and if I understood right the
change to concatenation may have coincided with starting to use separate
controllers, which will of course be a benefit in any case.
I will go with striping, using the default cylinder-size interlace.
Thanks to:
jerry@strobe.ATC.Olivetti.Com (Jerry Aguirre)
Kean Stump <kean@cerridwen.UCS.ORST.EDU>
billd@billd.cray.com (Bill Davidson)
glenn@cogito.usask.ca (Glenn Hollinger)
turboman@csn.org (Bill Miller)
len@contec.COM (Leonard Mills)
--Per Hedeland
per@erix.ericsson.se or
per%erix.ericsson.se@sunic.sunet.se or
...uunet!erix.ericsson.se!per
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:09:02 CDT