topic Re: EMC queue depth in Disk Enclosures

EMC queue depth

George A Bodnar — Thu, 22 May 2003 11:11:51 GMT

Does anyone have a good formula for what the queue_depth paramter should be set to (or if it should even be increased) for devices on this unit?

We have a mixture of meta-volumes (striped disks) and regular devices which by default all are getting a queue_depth of 8.

My hunch is the metas (4 regular devices) should probably be at least 16 but possibly as high as 32.

We are going to experiment on this, but was curious to see what others have done. A search of the forums and EMC support site shows generic references to queue_depth but nothing concrete.

Re: EMC queue depth

harry d brown jr — Thu, 22 May 2003 11:23:04 GMT

Do you have access to login into

http://powerlink.emc.com ?

If not, register and get a login, btw your emc rep can help you.

live free or die
harry

Re: EMC queue depth

Ross Zubritski — Thu, 22 May 2003 11:51:45 GMT

Depending the size of the data files as well as the DB block sizes, I would recommend either 16 or 32.

Regards,

RZ

Re: EMC queue depth

George A Bodnar — Thu, 22 May 2003 11:57:43 GMT

Yes have access to Powerlink - didn't find any interesting documents. We actually are working with our EMC rep on this and they gave us some documents - but once again nothing specific.

Just looking for some other opinions to see what other people are doing.

We are probably going to bump some of our meta devices to 16 and see if there is any change - good or bad.

Re: EMC queue depth

Stuart Abramson_2 — Thu, 22 May 2003 18:28:59 GMT

What is the "queue_depth" parameter?

What is it a parameter of? Not HP-UX?

Re: EMC queue depth

George A Bodnar — Thu, 22 May 2003 18:51:36 GMT

The queue_depth is set per device using the scsictl command. Also there are some kernel parameters (scsi_max_qdepth) and something else on 11i that affect this behavior for all devices.

Re: EMC queue depth

Tim Sanko — Fri, 23 May 2003 14:08:06 GMT

George,

Our queue depth is set to 16, according to our documentation on our setup documents.

We use 64GB Metas as well as 8.43 gb hypers for our HFS Filesystems. We trust the SYMM to protect our data.

Tim

Re: EMC queue depth

Tim Nelson — Thu, 28 Oct 2004 14:38:46 GMT

Hey George,
After a year has gone by did upping the queue depths ever have any good effects ?

Re: EMC queue depth

George A Bodnar — Wed, 03 Nov 2004 11:02:03 GMT

Honestly never really got a good feel for this - we actually backed out making changes due to other things that were in the queue.

I never found anybody (EMC/HP/other) that could state if/how this would help. In fact found a few sources that said they didn't believe it would make a difference.

Re: EMC queue depth

Alzhy — Wed, 03 Nov 2004 11:19:56 GMT

On our VxVM managed Solaris systems, I used to have an EMC hooked up and we left our Qdepth at 8 and maxphysIO at 1MB.

Someone posted this whitepaper on Q depth a few days ago. It might help.

As a general rule, Array Controller Port Qdepth => no.of LUNs x Q depth.

Re: EMC queue depth

Tim Nelson — Wed, 03 Nov 2004 11:26:49 GMT

Interesting Doc..
Any idea of what the effects to an application or system if the default queue depth was 8 ?
Is this a system throttle and once >8 application IO is told to wait but not error ?
I see the other extreme in the doc if set too high then the Array would return an error.
Is this analagy correct ?

Re: EMC queue depth

Alzhy — Wed, 03 Nov 2004 13:55:04 GMT

Tim..

Yes your analogy would be correct. And yes .. coupled with your Volume Manager (LVM+VxVM) and other SCSI tunables - Q Depth can be a system I/O throttle of sorts.

"High" Q Depth Settings (in my standards q depths greater than 16) mostly apply to environments that adopt very large LUN size standards and adopt a policy of a Filesystem or Server Volume Unit per LUN.

For XP 1024 (or XP12000)? The per port maximum is I think still 1024. That is why for this kind of array - I always stripe accross LUNs presented on different front-end ports and still bump up q depth from the 11i default of 16. Hitachi suggested staying at 8 on our 9960 (XP512 equivalent)..