Queue Full Condition

oh, have not see this one before. How is it different from IOmeter ?

Are you looking for a disk stress test tool? Something like this:

http://blogs.msdn.com/storwdk/pages/disk-stress-test.aspx

Actually what i'm trying to do is to cause an IOs to arrive at a storage controller port faster than they can be handled and to fill up the port queue. So, if the queue depth is exceeded, the storage system returns a queue full status to the host in response to an IO.

Do you run the risk of filling the HBA queues first? There are a lot of HBA settings that might affect your testing.

glen

i see following information in my knowledge base. May be use the metalun concept for the queue full condition..

"The HBA controls the queue depth for each LUN.  With a queue
depth of eight and two HBAs the server can drive 16 concurrent I/Os to
each logical volume.  With four HBAs and the queue depth still at eight,
the server can drive thirty-two concurrent I/Os.

This is critical when using EMC meta volumes.  With an eight-way
meta, the host only "Sees" a single LUN and will limit the queue depth
to eight"

that context was with HP-UX and default queue depth is 8 (scsi_max_qdepth) at the server side. Most of the time it is not altered

Hi, SKT

Thanks

Can you please guide me to this knoledge base post. I'm interested in EMC recomendations about defining the queue depth parameters. Don't you thinc queue depth of eight might be too low and cause QUEUE FULL conditions?

Hi, Kelleg.

Thanks. Can you throw me some tips?

Alec

Alec,

Each HBA vendor has a sectiom on their web site for EMC specific HBA's and drivers - there is an installation and config manual for each driver with the recommended settings for the HBAs. You might find something there.

I'll ask one of the performance engineers if they have a way to cause queue full on the array. Just an FYI, I've not seen that before.

glen

Try this:

Configure a 1+1 R10 and hit it from 2 Emulex HBAs with the default 32 threads per lun (iometer set to 64 threads for a random test, for instance).  Or a single Qlogic with default 256 threads total for the card.

glen

Thanks, i'll try to work on it.

Alec

A slight correction to that that last message - I asked engineering about this and this is their reply:

-------------------------------------------------------------------------------------------
A Raid1 1+1 will Qfull at 14+32 (number of data disks * 14 + 32)
So the R1 should qfull at 46+1.

Each path to the R1 can allow that number before qfulling so you need to have that number of commands (threads) arriving concurrently to the LUN via a single path and not two.  If you do have MPIO then you will need to boost the test to at least 46*2+1 = 93 threads.

A Emulex HBA with default EMC driver settings will never QFULL.  The QLOGIC HBA with default EMC driver  settings on Windows hosts will QFULL.

The ports will qfull in theory at 2048 commands.  More difficult to get this setup as you will need multiple hosts and tests have to run commands concurrently.

-----------------------------------------------------------------------------------------

glen

> The ports will qfull in theory at 2048 commands.  More difficult to get this setup as you will need multiple hosts and tests have to run commands concurrently

Does anyone know at what value a DMX port gets a "queue full" ?

sounds like they already said it is 2048.