NFS over TCP might be slower. I have had several customer issues where they were seeing many "NFS Server Not Responding" messages on their client, and remounting via UDP fixed it.
It turns out that most of these customers were on UDP previously, and somehow remounted (via a reboot without appropriate vfstab file) and the TCP default could not handle what the UDP had done.
I also just did a quick "mkfile" test from Sun to Sun. I had a /udp and a /tcp mount. mkfile on 10 meg files was about 5% faster on UDP. That was from a Solaris 2.5.1 client to a Solaris 2.5 server. No Network Appliance file server was involved. You should try this yourself. I simply did "time mkfile 10M foo" each time. The numbers were quite consistent.
I don't have any conclusions from this. Try experimenting with it and note performance differences. Depending on your network, the TCP could be better, but I tend to be skeptical about it, and if a customer can afford UDP over TCP I usually recommend it.
Chris Caputo wrote:
On Wed, 11 Jun 1997, Jennifer Dawn Myers wrote:
e.g., no snapshots, increase the number of inodes, no atime update, disable NFS over TCP, ...?
We are currently using NFS over TCP for news purposes. Is that a bad thing?
We are currently just using the NetApp box for some binaries groups. Our settings are:
===== OPTIONS ===== autosupport.doit WEEKLY_LOG autosupport.enable on autosupport.mailhost mailhost autosupport.noteto ccaputo@alt.net autosupport.to ccaputo@alt.net autosupport.from autosupport df_2gb_lim off httpd.admin.enable on httpd.enable off httpd.log.max_file_size 2147483647 httpd.rootdir XXX httpd.timeout 900 httpd.timewait.enable off ip.path_mtu_discovery.enable on minra off mount_rootonly on nfs.big_endianize_fileid off nfs.per_client_stats.enable off nfs.tcp.enable on nfs.v3.enable on no_atime_update on nosnap off nosnapdir off pcnfsd.enable off raidtimeout 4294967295 raid.reconstruct_speed 4 raid.scrub.enable on root_only_chown on telnet.hosts * wafl.maxdirsize 10240
===== SNAP-SCHED ===== 0 0 0 0
For security reasons our NetApp is assigned to private address space (10.x.x.x, RFC1918) so it can't been seen from the rest of the net. If you are exposed to the net, you might to do that. Any client machines would also need to be using 10.x.x.x addresses in addition to their normal reachable addresses.
Our raidtimeout is set to max since we don't ever want the machine to stop running, after it loses a disk. We'll just get it a new disk quickly. Having it stop is not an option.
To NetApp - Make these things cheaper so we can buy more. Their really good, but a dollar a meg is too high when the non-raid market is at $0.1 per meg. Also, make it so those extra PCI slots can be used for more SCSI cards, even on the low end machines. When you can't do things like this it really makes the company seem like it is driven by marketing droids obsessed only with competing with Auspex.
Chris