On 2008-01-22 23:22:01 +0000, andrew@cucumber.demon.co.uk (Andrew Gabriel) said:
In article <vilain-9D752E.00352422012008@comcast.dca.giganews.com>,
Michael Vilain <vilain@NOspamcop.net> writes:
Yes, you understand me correctly. Either AIX or Solaris should be the
NIS master or slave systems AND don't use Linux as a NFS fileserver.
Linux as an NFS fileserver -- agreed, just don't go there.
I've heard plenty of anecdotes about Linux being a terrible NFS server but I haven't really found any good list of the actual problems encountered in the various Linux versions. Is such a thing around, or does everyone "just know" that Linux can't serve NFS well?
And does "well" mean fast, correct, reliable, compatible, or something else?
Reginald Beardsley 23 January 2008 18:35:40 [ permanent link ]
Chris Ridd wrote:
On 2008-01-22 23:22:01 +0000, andrew@cucumber.demon.co.uk (Andrew
Gabriel) said:
In article <vilain-9D752E.00352422012008@comcast.dca.giganews.com>,
Michael Vilain <vilain@NOspamcop.net> writes:
Yes, you understand me correctly. Either AIX or Solaris should be the
NIS master or slave systems AND don't use Linux as a NFS fileserver.
Linux as an NFS fileserver -- agreed, just don't go there.
I've heard plenty of anecdotes about Linux being a terrible NFS server
but I haven't really found any good list of the actual problems
encountered in the various Linux versions. Is such a thing around, or
does everyone "just know" that Linux can't serve NFS well?
And does "well" mean fast, correct, reliable, compatible, or something
else?
Cheers,
Chris
A friend of mine told me a real horror story about Linux based NFS in a seismic processing shop. They'd set the largest documented block size for transfers and would experience weird failures under load. After several months of this someone dug into the kernel source. As I recall him telling me, the kernel source comment accompanying setting the larger block size was:
// I hope no one ever selects this
Seems it didn't reliably transfer the entire block because the logic needed had never been fully implemented. Hopefully it is now, but it's certainly typical of my experience w/ Linux.
On Thu, 2008-01-24 at 07:56 +0000, Chris Ridd wrote:
On 2008-01-23 20:53:28 +0000, Chris Cox <notccox@notairmail.net> said:
...
We have had ZERO problems with our SUSE based NFS servers. We serve
terabytes of disk and have been doing so for hundreds of users on
hundreds of clients, even loaded for more than 5 years.
Were those clients also Linux, or a mix of things?
We have a mixture of clients:
AIX 4.3.3, 5.1, 5.2 5.3 HPUX 11.00, 11.11 (11i), 11.23 (11iv2) (PARISC and Itanium) Red Hat 7.1, 7.3, 8.0, 9 RHEL 2, 3, 4, 5, 5.1 SUSE 8.2, 9.0, 9.2, 9.3, 10, 10.1, 10.2, 10.3 SLES 8, 9, 10 (both SLES and RHEL are supported 32bit and 64bit as well as IBM zSeries under zVM) Solaris 2.6, 7, 8, 9, 10 (x86 and SPARC) SCO Unixware 7.1.4 NCR SVR4 (had to make it an NIS slave, old style NIS) OSF/1 (Tru64 or Digital Unix) 4.0, 5.1
We are a software ISV. I've been administrating and porting code across all flavors since about 1983.
NIS servers run off of two SLES 10 boxes. Primary NFS servers serve up terabytes of shared disk to all of the clients listed above (including the home directories currently split across two growable 800G areas each). Same storage areas are also made available via SMB, ftp and http protocols. We converted away from direct attached SCSI to fibre SAN (some 2G some 4G) for storage this year and increased our storage from 12TB to about 65TB (not all of that is allocated today). Because of that conversion, our NIS/NFS servers are only showing 116 days of uptime. Prior to that, the servers had been up for over a year (we are supposed to have a mandatory outage each year, it didn't happen in 2007).
Network is a Cisco based 1Gbit to the host with 10GigE inter switch delivering full 1Gbit (2Gbit) bandwidth host to host. However, this infrastructure existed even back when we were 100Mbit.
Obviously, we do have some high risk clients, especially those that can only do NFS via udp. Because those platforms are not heavy users, we haven't seen any issue yet and eventually those platforms will die off (much like the fact that we finally killed HPUX 10.0 and Solaris 2.5.1 ... for example).