NetBackup and Gigabit 3

[cstorm]

I recently switched my media server from using HP's Auto Port Aggregation with 4 100mbs connections to one single Gigabit connection. I don't believe I am getting the throughput I should. Output from the bperror command shows this for before and after throughput readings:

with four 100mbs connections:
42 of 46 multiplexed backups total 83746180 at 6307.776
35 of 39 multiplexed backups total 24269049 at 1303.26
51 of 58 multiplexed backups total 79370774 at 4268.697
61 of 68 multiplexed backups total 77291846 at 3473.198
47 of 62 multiplexed backups total 116213509 at 5001.841
38 of 59 multiplexed backups total 95937215 at 4038.711
Average Throughput: 4065.5805

with one 1GB connection:
55 of 56 multiplexed backups total 20009270 at 240.796
35 of 40 multiplexed backups total 165499472at 3943.742
40 of 43 multiplexed backups total 45570198 at 1057.325
54 of 60 multiplexed backups total 22690859 at 421.943
39 of 46 multiplexed backups total 30367031 at 461.8
42 of 49 multiplexed backups total 30966456 at 409.528
Average Throughput: 1258.8676

What could cause the slow down?

 

[RenC]

cstorm,

We have recently started our performance testing on a private gigabit ethernet for our backups. During our final phase of testing we noticed significant performance slow downs when we ran a backup from multiple hp systems simultaneously. We did not see this from any other platform. We did major testing and have isolated it down to an HP issue. We saw this performance hit even with ftp and ttcp process, so we have completely removed NBU from the picture. We placed a call to HP a few weeks ago and they have it escalated within their group. It is interesting to find someone else having the same issue. I would urge you to open a call with HP as well.

 

[systemengineertechie]

Has there been any follow up from talking to HP concerning the performance slowdown?

 

[Stumpr]

Yes, I would like to know the response from HP.
Also, I would like to know if the tuning guideline was followed for HP-UX kernel tuning

http://seer.support.veritas.com/docs/243461.htm

Bob Stump
Incorrigible punster -- Do not incorrige

 

[Vela]

The HP kernel tuning does work very well.

We had very slow performance on the master itself and with tuning we were able to speed up backups to less than half their runtimes and the master was much more responsive when tyring to use the NB gui


Ryan

 

[jdespres]

I know with Solaris.....

gige works when it's set to 1000FD with auto set to ON....

and 100bt works better when set to 100FD Auto off...

 

[creakyjoe]

its not always that straightforward. for instance sun and cisco don't always play ball. ie the notorious autonegotiation. we never really won the performance battle with our suns/cisco. got as far as full duplex/no autoneg, and pauserx/pausetx but just left it at that. we get more pressing issues with slowdowns on our SAN than the backups as backups are never a business priority

Rich

 

[cstorm]

We saw increased performance when we added another Fibre Channel HBA to our HP server and rezoned our Fibre switch so that the tape drive load was better balanced.

 

[enterprisebu]

I was rechecking our *BUFFER* settings files and talking with Veritas when I was told to remember that my NBU settings are bound by my OS settings. I checked my UNIX tcp_sendspace and tcp_recvspace settings and found them at the default of 16k. Once I increased them all to 32k the backups really picked up in speed. It took awhile to figure out why our windows media server was out performing our gigE UNIX server.
They'd initially tried telling us it was the auto-neg but I'd heard from others already that the gigE and 10/100/1000 cards hadn't shown the same trouble with that setting as the 10/100 cards have.

 

[sm42]

We are having similar problems on a Solaris 9 box using Gigabit fibre.

The backup takes a long time to complete.

Is there a Veritas doc which provides tips on how to tune Solaris?

 

[Stumpr]

Minimum system requirements for the Solaris kernel when used with VERITAS NetBackup (tm), defined in /etc/system

http://seer.support.veritas.com/docs/238063.htm

Shared memory error

http://seer.support.veritas.com/docs/268122.htm

Bob Stump
Incorrigible punster -- Do not incorrige

 

[sm42]

Thanks for the quick response.

 

[sm42]

Forgot to ask, whilst using the Soalris netbackup client, does anyone change the default tcp parmeters?