Server Throughput during backup

[Dave Lloyd]

Wasn't sure of the appropriate area to post this but..
I'm setting up a new backup environment based around NetBackup 4.5 and have been very disappointed with the general throughput I'm getting.
I have 2 DL80's G3, 1 as master attached to an StorageTek L700 with 2 LTO2 drives (FC) and 1 as media server attached to a Quantum DX30 (FC).
Using either a DL380G3 or a ML530G2 I seem unable to get any more than about 12MB/s from the client during backup. The ML has 7 72GB drives (10k rpm) on a 5312 controller so in theory the Network should be the bottleneck - which is 'only' 1Gb.
I've tried bypassing the backup software by using FTP and get even worse throughput - arounf 9MB/s. Even configuring a 1.5GB disk in RAM and backing that up only gave about 12MB/s too.
I'm running Windows 2000 SP4 throughout, but have tried Solaris 8 clients (and ftp) with similar results.
I'm sarting to wonder whether there is a limit either on the PCI bus (but PCI-X ought to be well quick) or the TCP stack....large packets has no effect though.
What kind of throughput does everyone else get? Not so much artificial benchmarks, but real life file transfers (I use perfmon to determine throughput from/server). I can't believe that Gb connectivity is only being 10-15% utilised per machine....I'd have hoped at least 30% and prefer 50% or so - am I so wrong?

Thanks

Dave

[Thilo Knoch]

Quite no solution for the gigabit ethernet problem, but some tips for netbackup:

If you have a look after the statistics in the CPQTTAC (or whatever is the name beginning with HP now)applet in the status bar and you find transmit/receive underruns you could tune the registry parameter
HKEY_LOCAL_MACHINE\SOFTWARE\Veritas\NetBackup\CurrentVersion\Config. The default is 32 (decimal). You could test 128 or even 256.
see:
http://www.zedat.fu-berlin.de/services/fab/man/4.5/TuningNetbackupOnNT-4.5.pdf
and
http://seer.support.veritas.com/docs/240887.htm
(Yes it is for 3.41 but ...)

(Only a workaround for faster backup)
But aside from real network throughput the "client compression" feature of netbackup CAN really speed up backups (if you do not back up compressed data).
The only problem could be the relatively high CPU load on the client machine during backup.
(Policy attribute "compression" in jnbSA)

Have you tried to bypass your network infrastructure by using a crossover cable between two clients or client and server?
May be the switch and clients/server do not work together as expected.

[Olivier Drouin]

Check your cables.
Dont use auto-detect on your LAN card. Force the speed/duplex to your wanted value.

[John Merrill]

We recently had a similar problem; turned out to be a mismatch in bandwidth/duplex settings between the NICs and the switch. Set everything to 1000 full manually, and maybe you'll have better luck. Cheers, J

[Colin_29]

Dave, Check out the following link: http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=lpg50460&locale=en_US

[Dave Lloyd]

I had a bit of a breakthrough in todays testing.
Firstly I tried bypassing the switch by using a cross over - now, this meant re-installing NetBackup as a master server on my media server since the 'proper' master is across a MAN link in a different building.
The throughput jumped to 20-30MB/s on a large file and around 13MB/s on small files - basically double what I had before.
Putting it back on the switch I the figures remained similar- so it seems to be related to the master server somehow....more tests to come (like try a different switch just to be sure) but I think I need to hassle Veritas now.
Having said that, it doesn't explain the poor FTP performance.
For the record, the NIC's are hard set to 1000 Full (can't do half on Gb anyway) and so are the switch ports.
Thanks
Dave

[Dave Lloyd]

Thanks for the link Colin. I'd already tried those tools from that page prior to posting here.
My Dx30 virtual tape did 126MB/s
My LTO2 drive did 35MB/s at 2:1 compression and 90MB/s at 3:1.
The 'PAT' gave me varying figures but the main file server client was 9MB/s quick test, around 30MB/s full test and about 70MB/s sequential test.
My personal PC with a single IDE drive does 22MB/s, 17MB/s and 49MB/s for the same tests so i still believe that the RAID5 set is not operating as it should.