Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!cs.utexas.edu!csd4.milw.wisc.edu!bbn!mit-eddie!uw-beaver!rice!sun-spots-request
From: ktk@spam.istc.sri.com (Katy Kislitzin)
Newsgroups: comp.sys.sun
Subject: swapping over nfs
Keywords: SunOS
Message-ID: <3616@kalliope.rice.edu>
Date: 1 Jun 89 01:15:44 GMT
Sender: usenet@rice.edu
Organization: Sun-Spots
Lines: 50
Approved: Sun-Spots@rice.edu
X-Sun-Spots-Digest: Volume 8, Issue 11, message 15 of 15

I am in the proccess of converting my 4/260 from 3.2 (Sparc) to 4.0.3.  As
this is the fifth or so time I've installed 4.X, I got to wondering...
How does creating a few large swap files interact with the Berkely fast
file system ?  

Assumptions:

/export/swap is a seperate file system

All swap files are created at once, just after mkfs and before
there is any other filesystem activity

/export/swap contains *only* client swap files

Questions:

My understanding is that given this scenerio, all disk blocks are
allocated when the swap files are created; after that they are merely
modified.  So what function does the 10% minfree serve in this case?  Will
my performance degrade if I fill it up to 111% (as is the case on a 3/180
I configured a while ago)? I have not noticed any problems on the 3/180,
but not all the client partitions are in use...

More generally, it seems to me that the usual parameters newfs passes to
mkfs may not be optimal in the case of a few large files whose sizes never
change.  It seems like allocating few inodes, setting the block size to be
huge and setting minfree to be very low ( zero? ) would improve disk usage
and performance.  Has anyone played with this?

+++++++++

These sorts of considerations lead me to wonder about how the client
swapping and the server filesystem buffering interact.  

Do blocks of the client's swap file live in the server's buffer cache
until they expire and get written back to disk?

Is swap space allocated in such a way that there is a reasonably high
likelihood that requested pages will still be in the server's buffer cache
?  (assuming a reasonable amount of paging, like on a 4 M 3/50 )

if the above is true, it seems like one could improve clients' performance
simply by adding more memory to the server and allocating a large number
of filesystem buffers. comments?

Thanks for your time.  If this has been discussed in a previous sunspots
or if anyone has pointers to papers or SUN documentation on this topic, I
would be interested. 

--KT (ktk@spam.istc.sri.com)