wcs) (06/05/90)
In article <1990May29.225644.876@tc.fluke.COM> pwl@tc.fluke.COM (Paul Lutt) writes:
]The configuration examined was a Sun-3/390 fileserver running SunOS
]4.0.3 with twenty SparcStation 1 clients. [ each 8-16MB RAM ]
[ over 9000 blocks total activity,
74% swap writes
6% swap reads
20% real work (10% binaries, 4% root, 6% directories)
This means swap write:read is 12:1 !!
]The swapping behavior is very interesting. It appears that the virtual
]memory system in SunOS 4.0.3 tries to write out memory pages long before
]there is a need to reclaim any pages. Each time a new program is
]started, there is a corresponding flurry of writes to the swap file.
]Given the synchronous nature of NFS writes, this swapping traffic
]adversely affects the performance of both the workstations and the
]fileserver.
One thing I've wondered about for a while is how much you
really need writes to be synchronous. Sure, real data needs
to be written, but server machines are a lot more reliable
these days, and UPS has gotten cheap enough to use for servers.
Perhaps it's time to consider "sloppy-mode NFS", which lets
the server cheat on write-acks. Rather than change the
communication protocols, change the implementation so that
some file systems can be declared "sloppy", and used for
swap, tmp, etc. Yes, you lose some of your immunity to
crashes, but very little. Around here, the main causes of
crashes are either software bugs or power failures, both of
which would make a server-crash generally lose the active
processes on the clients anyway. So losing swap should be
no big deal.
--
Thanks; Bill
# Bill Stewart AT&T Bell Labs 4M312 Holmdel NJ 201-949-0705 erebus.att.com!wcs
# Actually, it's *two* drummers, and we're not marching, we're *dancing*.
# But that's the general idea.allyn@sdd.hp.com (Allyn Fratkin) (06/17/90)
In article <1990Jun4.230343.2796@cbnewsh.att.com>, wcs@cbnewsh.att.com (Bill Stewart 201-949-0705 erebus.att.com!wcs) writes: > One thing I've wondered about for a while is how much you > really need writes to be synchronous. ... > > Perhaps it's time to consider "sloppy-mode NFS", which lets > the server cheat on write-acks. hp-ux already has this option for nfs on hp-ux 6.2 (i think) and later. in the /etc/exports file, there is a -async option that can be applied to any file systems. this causes the server to immediately ack write requests but write the data asynchronously. we use it all the time and it works great. -- From the virtual mind of Allyn Fratkin allyn@sdd.hp.com San Diego Division - or - Hewlett-Packard Company uunet!ucsd!hp-sdd!allyn