NFS causing slow performance in cluster

[Krunoslav Kovacic]

Hello!
I have two node cluster (HP9000 K-class) each running different database instance (not OPS configuration!) with HP SureStore XP256 Disk Array. One volume group needs to be shared between cluster nodes meaning both nodes should be able to read/write to that volume group. This has been implemented with NFS, but unfortunatly it results in very slow performance. Is there a way to implement this functionality without NFS?

Thanks in advance,
Krunoslav Kovacic

-----------------------------------------------
oracle server 7.3.4
hpux 10.20

[Pramod_4]

Hi,

I don?t think that there is way of sharing a "file system" without using NFS in MC/SG.

But to improve the speed you may have to consider the following:-

Use HP MC/SG NFS tool kit and configure your NFS alone as a package with a floating IP address ( dedicate a private LAN between MC/SG nodes for NFS ) and put your NFS files system on a shared Disk drive.

Let the MC/SG mount the file system on the primary system.

On the secondary system, mount the NFS file system by using floating IP/host name so that the NFS traffic doesn't go through your main network.

Regards,

Pramod

[Krunoslav Kovacic]

Hi,

First of all, thank you for your answer.
Secondly, that has already been done, but results are unacceptable.

Here are some figures: 1.39 Mb/sec transfer rate when 100Mb file is copied to volume mounted "locally" by one node; 13Mb/sec when file is copied to the same volume mounted through NFS by another node.

Regards, Krunoslav

[JACQUET]

Hello,

NFS performances Pb depends on various things...
I can give you some ideas :
- First, try to know the Load Average of your server using NFS. If it isn't too much, you can increase the number of nfsd daemons (i.e modify /etc/rc.config.d/nfsconf). This will increase IO request and RPC calls.
- Second, check the buffer pages used by the system (with vmstat). Also if there is still free memory when all your servers processes are running, try to modify kernel parameters like nbuf, bufpages, or better use dynamic allocation bufpages by setting : bufpages=0, nbuf=0, and dbc_min_pct=5, dbc_max_pct=20.
Be careful of dbc_max_pct, it is using a percentage of total memory, ie : if you've got 1 Gb RAM, setting this param 20 results of allocating 200Mo for buffer cache...

It may help...
PJA

[Pramod_4]

Hi,

First of all it seems that you have to have heavy data transfer over the wire ( NFS ).

Look it to the tunable parameters pertaining to NFS ( BIOD and nfsd ) to obtain the maximum through put with the existing configuration.

To achieve your required throughput I would suggest you to use multiple Ethernet card with HP "Auto Port Aggregation ( APA )" product which would give you a cumulative data transfer throughput.

For Eg:-
If you have 4 x 100Mbps Ethernet card, With HP APA it could work as a 400 Mbps pipe between systems.

For more information about HP APA visit: http://docs.hp.com/hpux/netcom/index.html#Auto%20Port%20Aggregation%20(APA)

Regards,

Pramod

[Dan Hetzel]

Hi,

NFS performance is a known issue and is solved by patch PHNE_22125. Check with 'swlist' if it's been installed.

You'll find all info at the following link (there are quite a lot of prerequisites).

Dan

http://europe-support2.external.hp.com/wpsl/bin/doc.pl/sid=d1b91a2c15573fae0a/screen=wpslCandidateList

[Krunoslav Kovacic]

Hi,

Ethernet link utilisation during file-copy testing was about 15% (measured with Fluke instrument), that implies that bandwith of Ethernet link is not problematic. 15% of 100Mb/sec corresponds to 1.39MB/sec which was effective file transfer rate.

Best regards, Krunoslav

[MARTINACHE]

There are 2 versions of NFS : v2 and v3.

nfsstat should give the numbers of calls for each version.

If you only have v2 calls you can try to use only v3 which had better performances. (search a post for nfs and v3, I remember that you need a patch on 10.20)

Regards,

Patrice.