P4500 Multi-site and SQL 2008

[TFoster]

We are in the midst of implementing a Multi-site P4500 (2site-4node) solution. This is connected to new Blade servers running SQL 2008 r2. We have unable to obtain suitable IO subsystem performance for SQL. Would anyone be able to share details of a successful similar configuration (i.e. Switch models, network config, etc.)? We are at the end of our rope.

Thanks in advance.

Tim Foster

[Steven Clementi]

Can you explain how things are currently configured / connected?

What type of switches are the P4500's connected to?

how are the blades connecting to the storage?

what about to the user network?


Steven

[BennyO]

How many IOPS are needed for your SQL DB's? Running the nodes in Hardware RAID 5 will deliver 6,800 IOPS, running in RAID 10 it will deliver 9,600 IOPS. This assumes the volume is network RAID 10.
On the iSCSI switching side you should be using flow control over jumbo frames. If your switches support it then flow control and jumbo frames.

[TFoster]

I think HP calls this a "Campus" config
Site A
2x2910al SAN dedicated switches - VLAN for SAN traffic
2x 2x1Gb trunked Switch to Switch
Nodes 1 & 2 - 2x1Gb to each switch
Jumbo and Flow enabled
SQL Server (Blade) to SAN - 6Gb Fibre

Site A to B - 2x1Gb Fibre

Site B
1x2910al
Nodes 1 and 2 - 2x1Gb to switch

This has been escalated up through HP and VAR, checked by Lefthand people, etc. On Friday we tried 3500yl switches to no avail.

We (not the VAR or HP) have isolated the problem to sequential write performance for the LOG volume. It's just not there. If we move the logs to a DAS disk (C:), we get good performance in SQL.

That's why I asked for an example of a successful configuration for SQL - I know this one is not working for us.

[BennyO]

How do you have the NICs configured on the P4500's at Site A? Do you have SAN iq 9 installed?
Open up the CMC go to the Performance monitor. Add statistics for IOPS total for all your nodes in the cluster. Watch the value, you want to make sure it is not exceeding 1700 IOPS per node for long periods of time. If the IOPS look good then as you are writing data to the log watch the network utilization at the switch level.