| System Tuning |
|
| TCP/IP Tuning |
Linux-1U.net/NFS Tuning NFS
nfs.SourceForge.net Optimizing NFS Performance
MIT.edu TCP/IP Tuning
Redhat.com TCP Tuning
netperf, iperf or NetPipe
- Transfer speed on a typical 100Mpbs generic linksys/netgear consumer class hub/switch
- 256MB memory + ata-100 disks
- P4G - P4-2.4G = 60Mbps
- P4G - Cel-2.0G = 50Mbps
- P4G - via epia-800 = 20Mbps
|
| Common NFS Problems |
SCO.com trouble shooting
SCO.com
- You should be able to run at 75% of your rate gigE network speeds
- NFS symptoms:
- it copies .. and than it "stalls" or stops
- RPC: Time out
- NFS server hostname not responding, still trying
- Possible Solutions
- Change these options ( rsize,wsize ) should be set to rsize=1024 and wsize=1024
because fast servers cause data overruns on the ethernet driver on slow client machine
- Upgrade your Kernel
- Upgrade your NFS daemons and ulitiies
- Change your Network topology
- Split off windoze machines from solaris/unix machines
- Split off NFS v2 from NFS v3 machines
- Buy better quality switches - cisco, baynetwork, foundstone, etc
- Buy better quality NIC cables - plennum
- Buy better quality NIC cards - intel, syskonnect
- Shorten the length of your NIC cables , 50' (?) is the maximum specs
- Split off slow 10Mbps machines and printers onto its own switch
- Split off 100Mbps machines onto its own switch
- Split off 1Gbps machines onto its own switch
- Split off 10Gbps machines onto its own switch
|
| NFS Tuning |
NFSv4.org
nfs.SourceForge.net NFS-Howto
nfs.SourceForge.net perfomance
SourceForge.net
SourceForge.net nfs-utils
nfs.SourceForge.net NFS FAQ
SunHelp.org
IBM.com
IBM.com
IBM.com
IBM.com chap 8
IBM.com chap 9
IBM.com chap 10
IBM.com nfs parameters
IBM.com
SCO.com trouble shooting
SCO.com
SCO.com performance
PSC.edu
Redhat.com Network Tuning
Lists.Debian.org
BillHarlan.com NFS Tuning for Clusters ( local copy )
- NFS Server daemons
- NFS Client daemons
- To See the NFS options used
- mount
- showmount
- cat /proc/mounts
- cat /proc/fs/nfs/exports
- cat /proc/net/rpc/nfsd
- hard vs soft mounts
- hard == waits forever till the other end comes back
- soft == give up after a while -- you risk losing that data
- intr
use intr to ^C out of any hung application during NFS problems
- no_root_squash
- allow the remote root on the other machine to write your root-owned files
- you implicitly are trusting "the otehr machine"
- rsize, wsize
- use larger buffers to transfer files FASTER
-o rsize=1024,wsize=1024
-o rsize=4096,wsize=4096
-o rsize=8192,wsize=8192
-o rsize=16384,wsize=16384
-o rsize=32768,wsize=32768 --> nfs v3 only
- sync vs async vs wdelay
- /etc/fstab
mail:/var/spool/mail /var/spool/mail nfs soft,intr,rw,rsize=8192,wsize=8192 0 0
- /etc/auto.master and /etc/auto.net
mail -fstype=nfs,rw mail:/var/spool/mail
home -fstype=nfs,rw home:/home
- /etc/modules.conf
alias eth0 eepro100
options eepro100 full_duplex=1
|
| NFS Threads |
- grep th /proc/net/rpc/nfsd
- vi /etc/rc.d/init.d/nfs ( /etc/sysconfig/nfs )
# increase number of threads from default of 8 to 16 nfsd
RPCNFSDCOUNT=16
# RPCNFSCOUNT=64
|
| NFS Buffers aka lost NFS packets |
ilug-bom.org.in
- increase the buffer size
- echo 524288 > /proc/sys/net/ipv4/ipfrag_low_thresh
- echo 524288 > /proc/sys/net/ipv4/ipfrag_high_thresh
- vi /etc/sysctl.conf
#
net.core.rmem_default = 262144
net.core.rmem_max = 262144
#
# Increase the fragmented packet queue length
net.ipv4.ipfrag_high_thresh = 524288
net.ipv4.ipfrag_low_thresh = 393216
|
|
