tuning: NFS/ADVFS

From: Wayne Thompson <wayne_at_hal.com>
Date: Tue, 06 Oct 1998 14:49:28 -0700 (PDT)

We have a 2-node 8200 running 4.0D PatchKit 2/Trucluster 1.5 that's
intended to provide dedicated, high perf/avail NFS service.

We were expecting *much* better numbers than we got and tuning only
hurt. Can anyone suggest how we can improve the performance (hopefully
dramatically :)

Wayne

The details:

Untuned, that max SpecSFS 2.0 numbers we could rack up were 1400 ops/s
using the standard mix. For a single exported fset and 1600 ops/s for
two fsets. When I tuned it, the two fset number dropped to 1400 (I
haven't tried the single fset run tuned.)

The systems each have 2GB main memory, 2x Alpha,vEV5_at_613Mhz, 4x HSZ 70.
The data disks, 144x 9GB, are grouped into 24 raidsets of 6 disks
and then cut into four partitions each. These are then grouped into
12 domains comprised of two devices from each of the four HSZ ctlrs
for eight devices per domain.

The NICs involved were FastEthernet tulips (half-duplex) over dedicated
segments.

The tunings:
vm:
        ubc-nfsloopback = 1
        vm-ubcbuffers = 1024
        vm-ubcdirtypercent = 100
        vnode_age = 600
        vnode-deallocation-enable = 0
        name-cache-hash-size = 32768
        buffer-hash-size = 32768
        inode-hash-size = 16384

proc:
        maxusers = 4096
        
vfs:
        sys-v-mode=0
        vnode-age=600
        vnode-deallocation-enable=0
        name-cache-hash-size=32768
        buffer-hash-size=32768
        bufcache=30
        
ufs:
        inode-hash-size=16384

advfs:
        AdvfsCacheMaxPercent=30
Received on Tue Oct 06 1998 - 21:50:27 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:38 NZDT