Path: utzoo!utgpu!watmath!clyde!att!osu-cis!tut.cis.ohio-state.edu!mailrus!cornell!uw-beaver!rice!sun-spots-request
From: eggers@cs.buffalo.edu (Bill Eggers)
Newsgroups: comp.sys.sun
Subject: tunefs and swap files
Message-ID: <8812161921.AA09682@sybil.cs.Buffalo.EDU>
Date: 29 Dec 88 21:34:37 GMT
Sender: usenet@rice.edu
Organization: Sun-Spots
Lines: 42
Approved: Sun-Spots@rice.edu
Original-Date: Fri, 16 Dec 88 14:21:57 est
X-Sun-Spots-Digest: Volume 7, Issue 77, message 6 of 14

Hello,

I've come across an interesting option in the tunefs(8) manual page in
4.0:

	-e maxbpg

This indicates the maximum number of blocks any single file can allocate
out of a cylinder group.  The manual page says that this number is
typically 1/4 of the total number of blocks contained in a cylinder group.
It says that this default setting causes long seeks to be performed on
large files which could otherwise be placed contiguously in the cylinder
groups.  It also says for large files, this parameter should be set
higher.  Has anybody tried to do this with swap files?

>From what I understand of 3.X, nd assigns each client's swap space onto
into their own parts of the disk, so the entire swap file is contiguous on
the disk.  However, in 4.0, the swap file is just like any other file, so
it is created according to the parameter of maxbpg stated above.  Thus, if
swap files were created with the above parameter set to 100% of the total
number of blocks contained in a cylinder group, then the seek time should
go down when swapping, and 3/50's possibly run faster when swapping takes
place.  According to my calculations, a cylinder group on our 327 meg SCSI
disks is 4 Meg, or 512 blocks (?), and with my 16 Meg swap files, the swap
files are big enough to consider tweaking this parameter.

On our systems, the swap files reside on partitions separate from non-swap
files, so the tunefs can be applied to those partition(s).  However, if
this isn't so on other systems, one could temporarily tunefs it, create
the swap files, then un-tunefs it.  

We don't have enough identical clients to compare the performance before
and after the tunefs.  Could somebody else perhaps try this out to see if
it improves (or degrades) the swapping time?  Or, Please tell me if this
is already common practice.  If there is some flaw in my reasoning, then
please let me know.

		Thanks.

Bill Eggers             Dept. of Chemistry, SUNY at Buffalo
103 Acheson Hall        Buffalo, NY 14214
eggers@cs.buffalo.edu