This maybe trivial, but try http://nfs.sourceforge.net/nfs-howto/performance.html
-----Original Message----- From: Gregory Gulik [mailto:greg@gulik.org] Sent: Wednesday, June 30, 2004 1:27 PM To: For users of Fedora Core releases Subject: Re: VERY slow NFS FC2 -> RH9
Ah, good point, but upon checking:
Client: eth0: VIA VT6102 Rhine-II at 0xe8800000, 00:0c:6e:ec:87:0f, IRQ 10. eth0: MII PHY found at address 1, status 0x786d advertising 01e1 Link 41e1. eth0: Setting full-duplex based on MII #1 link partner capability of 41e1.
Server: eth0: Intel(R) PRO/1000 Network Connection e1000: eth0 NIC Link is Up 100 Mbps Full Duplex
I checked both switches and those links are both being indicated as full-duplex.
Here's another data point. I used scp to copy the same 11MB file that that took 1.7 seconds. That's really pointing at an NFS problem rather than a network problem.
Yang Xiao wrote:
Hi, Are you sure they are all running at 100/FULL? Try set both
interfaces
to 100/FULL instead of AUTO. Yang
-- Greg Gulik http://www.gulik.org/greg/ greg @ gulik.org
-- fedora-list mailing list fedora-list@redhat.com To unsubscribe: http://www.redhat.com/mailman/listinfo/fedora-list
Actually, I think I got it. I removed the "sync" from the options and it sped up immediately. The copy now took 2 seconds.
Hmm, I've been using that option for quite a while, probably at least since this client was on RH9 but only upon upgrading to FC2 did I find NFS performance to be this poor.
At least the problem is solved. Thanks!
Yang Xiao wrote:
This maybe trivial, but try http://nfs.sourceforge.net/nfs-howto/performance.html
On Wed, Jun 30, 2004 at 12:47:07PM -0500, Gregory Gulik wrote:
Date: Wed, 30 Jun 2004 12:47:07 -0500 From: Gregory Gulik greg@gulik.org To: For users of Fedora Core releases fedora-list@redhat.com Subject: Re: VERY slow NFS FC2 -> RH9 Reply-To: For users of Fedora Core releases fedora-list@redhat.com
Actually, I think I got it. I removed the "sync" from the options and it sped up immediately. The copy now took 2 seconds.
Hmm, I've been using that option for quite a while, probably at least since this client was on RH9 but only upon upgrading to FC2 did I find NFS performance to be this poor.
At least the problem is solved. Thanks!
Well make sure to make a note in your system notebook about turning off the sync option.
I have mixed opinions on the sync bit. I have never found data corruption issues that matter with modern uptimes but toggling this bit is one of the classic benchmark tricks. If specific data integrity requirements justifies then you will want to set it back. Some data base files need it.
Other tuning tricks have to do with matching read and write sizes to the data chunkieness. Also file locking...
I think you had a Rhine controller. I have problems with my cheap switch and this controller in full duplex. Switching to half duplex solved my packet loss and speeds then ran at 100BaseT(HD) speeds. If some of the NFS sync reply packets were getting lost then things slow down a lot. For me scp was ok, but rsync was not when I was running in full duplex mode.
Anyhow when you see retransmit packets in tethereal or tcpdump output between local systems switch from FD to HD and see if it helps.