ISCSI Target under Linux - Performance?

[Krobar]

Hi All,


I'm quite late into the process of build and config for a pair of ISCSI servers at work and need some help with expected performance.

Hardware is already bought and the drives arrive next week so I can start performance work. ISCSI array is a 10 disk Raid 6 per server (7200RPM SAS) on an integrated LSI 9265 with BBU. The servers will replicate using DRBD (Over Infiniband/SDP). Functionally the servers are up and running on 1 test drive each at the moment so all that is left is performance and stress testing once the drives arrive.

I know where things should be with seqential performance but have no idea what sort of random/IOPS performance to expect. Can anyone offer up some IOMeter or random perfornmance results for an ISCSI client machine on a 6-12 disk array? Any tips with optimising performance of a TGT target?

 

[msitpro]

Try this for a rough idea:

http://www.wmarow.com/strcalc/

I don't think it'll be pretty for IOPS on datasets larger than the RAID controller's cache. What's it due to be used for?

 

[Krobar]

Try this for a rough idea:

http://www.wmarow.com/strcalc/

I don't think it'll be pretty for IOPS on datasets larger than the RAID controller's cache. What's it due to be used for?

Thanks,

Disks we are using are 2.5" constellation which looks like identical performance to the listed 3.5" version. Looks like 220 IOPS is what to expect.

The usage is for storing iscsi volumes for a number of test VMs. We are moving from 8 servers with a single disk each to them all being fed by ISCSI because of data security, redundancy and backup reasons (Extra performance is not expected although hopefully it wont be terrible). Any tips?

 

[Krobar]

The onboard controller being used is an LSI 9265 as in the review below:
http://www.tomshardware.com/reviews/sas-6gb-raid-controller,3028-12.html

The IOMeter results on the Apache and DB workloads which most closesly mimic our workload do not show anywhere near the performance dropoff for Raid6 suggested by the calculator, I guess the calculator was based on old raid card results.

Using the workstation pattern because its easiest to work out here is the review results vs calculator:
Raid 0: 1782 / 1583
Raid 5: 1397 / 904
Raid 6: 1252 / 703

Adjusting calculated to real I get about 582 IOPs for a workstation profile on the 10disk 7200rpm setup. More interested now as to what the real results will be

 

[msitpro]

The onboard controller being used is an LSI 9265 as in the review below:
http://www.tomshardware.com/reviews/sas-6gb-raid-controller,3028-12.html

The IOMeter results on the Apache and DB workloads which most closesly mimic our workload do not show anywhere near the performance dropoff for Raid6 suggested by the calculator, I guess the calculator was based on old raid card results.

Using the workstation pattern because its easiest to work out here is the review results vs calculator:
Raid 0: 1782 / 1583
Raid 5: 1397 / 904
Raid 6: 1252 / 703

Adjusting calculated to real I get about 582 IOPs for a workstation profile on the 10disk 7200rpm setup. More interested now as to what the real results will be

It all depends how much of your workload hits the cache of the controller....

Writes will be pretty quick....until you have a disk fail and the cache gets disabled.

When this happens you will have a perfect storm of:

1. RAID 6 is dog slow for writes in a healthy state.
2. The array will be in degraded state and rebuilding data onto a replacement drive.
3. The RAID controller will likely switch off it's cache during the rebuild process.
4. Rebuilding a 10x 1TB disk RAID 6 could take several days.


...although you say they are test VMs so it may not be a problem.