Friday, September 08, 2006

HP EVA8000 Autopath Tuning

Here is some information that might be useful to you. Following our RAC cluster setup and install, the EVA 8000 performance was very disappointing. First, our Oracle performance was disturbing. Then we did some more I/O benchmarks using "dd" and found that I/O performance was actually worse than the first baseline benchmark. So, this indicated a problem with the underlying storage and not the Oracle setup.

The UNIX administrators, HP, and my peers began looking into the problem. I wasn't totally engaged but want to post the information anyway. Props to my colleages. This is more system administration territory; however, as DBAs we need to know the impact of EVA tuning.

HP checked the load balancing policy on the nodes. It was set to “No Load Balancing” which greatly impacts performance. Now, how it was changed to "No Load Balancing" after the first I/O benchmarks is still a mystery.

To see the LUNs, issue the command “autopath display” as root. This will list all the LUNs and show the HP Autopath load balancing policy.

root@hostname:/ # autopath display
...
Load Balancing Policy : No Load Balancing
...

So, HP and my peers recommended that the HP Autopath load balancing policy be set to round robin for all LUNs. The autopath set_lbpolicy command sets the load balancing policy for the specified device path.

autopath set_lbpolicy <{policy name} {path}>
description: sets load balancing policy
usage: autopath set_lbpolicy <{policy name} {path}>
Policy name: The load balancing policy to set
Valid policies are
RR : Round Robin.
SST : Shortest Service Time.
SQL : Shortest Queue Length.
NLB/OFF : No load Balancing.
Path : Device Special File e.g./dev/dsk/c#t#d#

Example:
# autopath set_lbpolicy RR /dev/dsk/c0t0d0
The example above sets the policy to Round Robin.

Here is a little more information about what the policies mean:
  • RR : Round Robin. - I/O is routed through the paths of the active controller in round-robin order
  • SST : Shortest Service Time. - is a measurement against the average service time of the IOs on a path
  • SQL : Shortest Queue Length. - is a calculation on the device queue depth when it is on a certain path.
  • NLB/OFF : No load Balancing. - No consideration given to service times or queue lengths, typically impacts performance.

4 comments:

Unknown said...

Thanks for your valuable information, We're implementing a HP-UX PA-RISC RAC with securepath. Are you using the Service Guard extention for RAC? Do we need it?

Best Regards

sedwardba said...

Hi Alfonso. No we are not using Service Guard. Oracle recommended for us not to use it and just rely on Oracle Clusterware to handle failover. SG is kind of expensive and it saved us some money :-)

Unknown said...

Thanks for your fast answer.
It would be possible for you tog ive me your e-mail address, I would like to ask you some questions. My e-mail is aleon68@gmail.com

Regards

sedwardba said...

Sure sending you an email now.