NFS Problems/Questions

Sat Jun 30 23:33:23 UTC 2007

On Sat, Jun 23, 2007 at 07:42:24PM -0400, Jason Morgan wrote:
> On Sat, Jun 23, 2007 at 12:46:27PM -0700, Michael Smith wrote:
> > Hello Jason:
> > 
> > On Jun 23, 2007, at 9:34 AM, Jason Morgan wrote:
> > 
> > >I've been having some trouble with NFS performance for some time and
> > >now that class is out, I've had a bit of time to investigate but I'm
> > >stuck. Below are the details of my investigation. Hopefully, someone
> > >here can give me some advice.
> > >
> > >The basic problem is that my NFS performance is very slow. Right now,
> > >I am connecting two workstations to a NFS server, which has my home
> > >directory, etc, mounted. They are connected over a gigabit network
> > >(right now with mtu set to 7000, which is supported by all hardware --
> > >changing it to 1500 has no effect on performance, which is
> > >strange). Each system is running 6.2-RELEASE or -STABLE. Each system
> > >is also using the following network card:
> > >
> > ># ifconfig sk0
> > >sk0: flags=8843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> mtu 7000
> > >        options=b<RXCSUM,TXCSUM,VLAN_MTU>
> > >        inet 10.0.0.2 netmask 0xffffff00 broadcast 10.0.0.255
> > >        ether 00:17:9a:bb:05:87
> > >        media: Ethernet autoselect (1000baseTX <full- 
> > >duplex,flag0,flag1>)
> > >        status: active
> > >
> > ># dmesg | grep sk
> > >skc0: <D-Link DGE-530T Gigabit Ethernet> port 0xec00-0xecff mem
> > >      0xfdff8000-0xfdffbfff irq 18 at device 10.0 on pci0
> > >skc0: DGE-530T Gigabit Ethernet Adapter rev. (0x9)
> > >sk0:  <Marvell Semiconductor, Inc. Yukon> on skc0
> > >sk0:  Ethernet address: 00:17:9a:XX:XX:XX
> > >
> > >## Server /etc/rc.conf settings
> > >
> > >rpcbind_enable="YES"
> > >rpc_lockd_enable="YES"
> > >rpc_statd_enable="YES"
> > >nfs_server_enable="YES"
> > >nfs_server_flags="-u -t -n 12"
> > >nfs_bufpackets="32"
> > >mountd_flags="-r"
> > >
> > >
> > >## Client /etc/rc.conf settings
> > >
> > >nfs_client_enable="YES"
> > >nfs_bufpackets="32"
> > >nfsiod_enable="YES"
> > >nfsiod_flags="-n 6"
> > >rpc_lockd_enable="YES"
> > >rpc_statd_enable="YES"
> > >rpcbind_enable="YES"
> > >
> > >## /etc/exports
> > >
> > >/usr -alldirs,maproot=root client1 client2
> > >
> > >
> > >For performance benchmarking, I am using dd. Locally from the server,
> > >this is a representative result when writing a 1GB file:
> > >
> > >## Local write test (for an upper-bound on what to expect).
> > >
> > ># dd if=/dev/zero of=./nfs.dat bs=1024k count=1000
> > >1000+0 records in
> > >1000+0 records out
> > >1048576000 bytes transferred in 19.580184 secs (53552919 bytes/sec)
> > >
> > >Connecting from a client (both clients get approximately the same
> > >results).
> > >
> > >## Remote connection (UDP), mounted in /etc/fstab as with flags:
> > >## rw,-U,-3,-r=32768,-w=32768
> > >
> > ># dd if=/dev/zero of=./nfs.dat bs=1024k count=1000
> > >1000+0 records in
> > >1000+0 records out
> > >1048576000 bytes transferred in 101.151139 secs (10366428 bytes/sec)
> > >
> > >## Remote connection (TCP), mounted in /etc/fstab as with flags:
> > >## rw,-T,-3,-r=32768,-w=32768
> > >
> > ># dd if=/dev/zero of=./nfs.dat bs=1024k count=1000
> > >1000+0 records in
> > >1000+0 records out
> > >1048576000 bytes transferred in 59.668585 secs (17573334 bytes/sec)
> > >
> > >As can be seen above, TCP is much faster than UPD. I have tried many
> > >different mount settings and these are the best results I could
> > >get. To test whether or not I have having network issues, I
> > >transferred the same nfs.dat file via a http connection and got
> > >~32MB/sec -- almost 2x the speed of the TCP NFS connection. 32MB/sec
> > >is about what I would expect given that my fastest write speed is
> > >~50MB/sec.
> > >
> > >At this point I am stumped. I have tried increasing/changing the
> > >number of nfsiod servers as well as nfs_bufpackets. No matter what
> > >settings I change, the results are always the same. I get only two
> > >errors, first on /var/log/messages on the server I have just begun
> > >seeing:
> > >
> > >Jun 22 21:13:47 crichton routed[666]: sendto(dc1, 224.0.0.2):  
> > >Operation not permitted
> > >Jun 22 21:13:47 crichton routed[666]: sendto(sk0, 224.0.0.2):  
> > >Operation not permitted
> > >Jun 22 21:13:50 crichton routed[666]: sendto(dc1, 224.0.0.2):  
> > >Operation not permitted
> > >Jun 22 21:13:50 crichton routed[666]: sendto(sk0, 224.0.0.2):  
> > >Operation not permitted
> > >
> > >This appeared after I added a route; however, I added the route after
> > >many of the tests were done. I get the same results now as before the
> > >new route. On one of the clients (the one running 6.2-RELEASE-p1), I
> > >also get a nasty error:
> > >
> > >nfs/tcp clnt: Error 60 reading socket, tearing down TCP connection
> > >
> > >This cropped up last night after I tweaked some settings. They have
> > >now been changed back, but I still get this error. The other client is
> > >unaffected.
> > >
> > >I appreciate any help people can provide on tracking down the
> > >issues. Sorry about the long email -- just trying to be thorough. Of
> > >course, I've searched the Internet and can't find any clear assistence
> > >on these issues.
> > >
> > >Cheers,
> > >~Jason
> > >
> > We use the following settings on a mail cluster that's pushing about  
> > 50 MB/sec sustained.
> > 
> > 10.211.1.213:/m0        /mail/m0        nfs      
> > rw,tcp,intr,noatime,nfsv3,-w=65536,-r=65536
> > 
> > # NFS Server
> > rpcbind_enable="YES"
> > rpc_lockd_enable="YES"
> > rpc_statd_enable="YES"
> > nfs_server_enable="YES"
> > nfs_server_flags="-u -t -n 16 -h 10.211.1.213"
> > mountd_flags="-r"
> > 
> > I would imagine the larger read/write values above would be fine for  
> > you as well, given you have Gigabit links.  The 'noatime' setting may  
> > be problematic depending on your application.  You might want to  
> > Google specifics on what applications use atime to see if that's a  
> > good flag for you or not.
> > 
> > I'd love to see your results if you decide to test those settings  
> > offline.
> > 
> > Regards,
> > 
> > Mike
> 
> Mike, thank you for the advice. I followed it, but didn't make much
> progress. I am still pushing the same ~17M/sec. I did mess with
> net.inet.tcp.sendspace and net.inet.tcp.recvspace some and managed to
> get a couple more MB/sec out of the connection, but not much
> else. It's quite possible I am missing something obvious.

I'm still completely stuck on this. I wasn't able to get any better
performance out of NFS from the suggestions above. Thought maybe it
was a problem with my 6.2-RELEASE machines (though I had no evidence
that it was). I decided to update all machines to -STABLE. As I
suspected, no luck. Anyone have any other suggestions?

Thanks,

~Jason