Unsolved

This post is more than 5 years old

3 Posts

67861

May 3rd, 2010 09:00

Poor performance - MD3000i as a file server

Greetings folks

I have an MD3000i acting as storage for a PE2950; below is my observation:
PE2950 file copy local drive to MD - ok throughput, up to 70% of GbE, about roughly the same when file copy is from MD to local drive
Client on the public LAN file copy from share on PE2950 (local drive), ok throughput also

However, when the client maps to the file share, and the share happens to be from the MD - I obviously see roughly equal utilisation on both the public LAN and iSCSI LAN NICS (both using the server onboard NICs); but performance is very bad like 7% of GbE.

First thing that came to mind was that the PE2950 couldn't drive both NICS at the same time, seeing that testing from client > server local disk or server local disk > MD worked ok; but the PE2950 is not ancient hardware either (4GB RAM, 2.x GHz), so it can't be that bad?

Thought it might be due to interrupt sharing, but it's been ages since I had to tweak that on a server, but anyway I've tried tweaking those but it does not seem to help.

Does anyone have any leads on where to go to? Firmware is latest on the MD3000i. (15 x 600GB 15K SAS), also updated the SAS disk firmware.

Thanks!

2 Intern

 • 

847 Posts

June 21st, 2010 12:00

These are identical MD3000i's? or is one an MD3000 and the other an MD3000i?

6 Posts

June 21st, 2010 14:00

Yes these are the same MD3000i. One of them has just one spare disk. I do some test again with IOMeter and i find that if i configure 2 workers to read 32K on the first MD3000i, i've got pretty good perf. To be sure, i've changed the Ethernet cable but no luck. Tomorrow, i'll try to make configuration from "scratch" except that i've some data on the virtual drive. For more information, my virtual drive is 2,7TB large mounted as GPT under Windows 2003 x64.
Really i've got headache because i really don't see any differences between my both equipments.

6 Posts

June 24th, 2010 07:00

Feedback:
Finally i've found a solution. I change the MTU on the Broadcom iSCSI Adapter (from 9000 to 8000). Even if the iSCSI port on the MD3000i is still configured with a MTU of 9000, reading performance really improve. Before i had poor 1 Mo/s throughput for reading now i've got almost 30Mo/s (and I/O increase from 30 to 1000 !!). Anyway, i'm still looking for what happened and why i need to change MTU.
Hope this can help.

June 24th, 2010 10:00

"Feedback:
Finally i've found a solution. I change the MTU on the Broadcom iSCSI Adapter (from 9000 to 8000). Even if the iSCSI port on the MD3000i is still configured with a MTU of 9000, reading performance really improve. Before i had poor 1 Mo/s throughput for reading now i've got almost 30Mo/s (and I/O increase from 30 to 1000 !!). Anyway, i'm still looking for what happened and why i need to change MTU.
Hope this can help.
"
I have seen this issue before and it is the switch causing it. Some switches (not all) when MTU is set to 9000 will only allow packets of exactly 9000 (or less) through without fragmenting them. Usually a NIC set to MTU of 9000 really sends packets slightly larger than 9000, which the switch then fragments and iSCI hates fragmented packets. This causes prformance to drop to the floor like you have seen If possible set the MTU on the switch to 9216 and if that isn't possible put the NIC MTU as high as you can that is still under 9000.

6 Posts

June 24th, 2010 15:00

You're right Aaraon, implementation of Jumbo Frames seems a bit different for some NIC and switch manufacturers. Anyway, in my case there's no switch between my server and my MD3000i ( direct link). Finally, i've configured NIC MTU to 8500 and perf are good enough.

2 Intern

 • 

847 Posts

June 28th, 2010 10:00

Does this san have single controller? Even of it doesn't check into this....

http://vmtoday.com/2009/06/ibm-ds3300-iscsi-write-performance-solved/


Basically it may be cache related.
No Events found!

Top