nfs performance between Red Hat 6 server and HP-UX11i clients

[romano r]

Hello,

I've a file server RHEL6 on a VMWare environment that shares via nfs its data to RHEL4,5,6 and HP-UX11i clients.

Since this machine will substitute the actual file server based on HP-UX11i I'm doing some performance test, trying to tune a little the new server.

One of the issue I encountered is the nfs performance of the hp-ux clients that is quite poor in respect to the linux ones.

I did some test in reading/writing files of the usual size we use and some test with the command referenced in different forums: "time dd if=/dev/zero of=/mnt/test/foo bs=16k count=16384", up to now I haven't used specific tools like iozone and so on.

The tests show me that on linux clients I reach the speed of about 100MB/s in write and 300MB/s in read, on hp-ux about 10MB/s on both mode.

I left the nfs to autonegotiate wsize and rsize parameters, they are: 524288 for Linux and 32768 for hp-ux, unfortunately it seems that they are already the max the kernel allows, or at least I do not know how to increase them, if possible, in the hp-ux kernel, I played increasing and lowering them with different values in the mount command, without any improvment of the performance.

I also tried to change from tcp to udp, nothing.

Has someone any suggestion?

here some data of the linux server:

Linux mybox.com 2.6.32-131.17.1.el6.x86_64 #1 SMP Thu Sep 29 10:24:25 EDT 2011 x86_64

/dev/mapper/vg00-lvol1
                     1008M  320M  638M  34% /
tmpfs                 1.5G     0  1.5G   0% /dev/shm
/dev/sda1              97M   76M   17M  83% /boot
/dev/mapper/vg00-lvol5
                     1008M   36M  922M   4% /home
/dev/mapper/vg00-lvol6
                      2.0G  197M  1.7G  11% /tmp
/dev/mapper/vg00-lvol3
                      3.9G  1.6G  2.1G  43% /usr
/dev/mapper/vg00-lvol4
                     1008M  434M  524M  46% /var
/dev/mapper/vg00-lvol7
                     1008M  155M  803M  17% /opt/mng_archives
/dev/mapper/vg10-lvolA
                      296G   15G  267G   6% /DYN_ARCHIVE

MemTotal:        3089344 kB
MemFree:         1642020 kB
Buffers:          387160 kB
Cached:           595488 kB
SwapCached:            0 kB

test results attached

Thank you to all

bye

Romano

[Dave Olker]

What OS are the 11i boxes running?  11.11?  11.23?  11.31?  Once I know which OS you're using I can offer suggestions to help improve performance.

Dave

[romano r]

hello, 11.11

tnx

[Dave Olker]

11.11?  That's unfortunate since we made so many improvements to the NFS client code in 11.23 and 11.31.  Any chance of updating to a newer HP-UX release on the client?

Even if you stay on 11.11, you should be able to get more than 10MB/sec.  It almost sounds like the 11.11 system is using a 100BT network interface and the other systems are using GigE.  Are you certain the 11.11 system is using GigE network interfaces for the NFS traffic?

Before diving into NFS, let's make sure your network connection is capable of pushing data at rates similar to the Linux system.  Once we ensure the networking is solid we can look at NFS. 

I'd suggest getting the latest copy of netperf (http://netperf.org) and compiling the latest version on both a Linux client and the 11.11 client.  Install the netperf server daemon (netserver) on the common NFS server.  Then use netperf on the client (Linux and 11.11) to see how fast they can send/receive data to the remote NFS server.  Begin with inbound and outbound streaming tests and then we can move to RR (request/response) since that more closely simulates how NFS works.

Sample netperf syntax to start with:

# netperf -t TCP_STREAM -H <IP address of NFS server> -f M -- -s 256K -S 256K -m 32K

# netperf -t TCP_MAERTS -H <IP address of NFS server> -f M -- -s 256K -S 256K -m 32K

Regards,

Dave

[romano r]

Hi Dave,

thank you a lot for the suggestions, actually I've to work on the network before, as you suggested I found a couple of problem that are limiting the speed of the lan:

1) some HP-ux client with 100MB/s lan network

2) some lan switch is limiting the bandwidth of other HP-ux servers.

and finally, we are moving away form HP-UX in favor of RHEL OS.

Bye

Romano