topic Re: XP256: PV timeout and SCSI queue depth in Disk Enclosures

XP256: PV timeout and SCSI queue depth

Mark DeBoer — Wed, 07 Feb 2001 20:53:38 GMT

The HP-UX config guide for the XP256 says that you should set the physical volume timeout to 60 seconds and the SCSI queue depth to 2 for each LUN on the XP256 with OPEN-3, OPEN-8, or
OPEN-9 emulation type. Is this true for other emulation types such as OPEN-E?

Re: XP256: PV timeout and SCSI queue depth

Melanie Battson — Wed, 07 Feb 2001 22:59:24 GMT

Interesting question, however, we are having this discussion right now with our friends at HP. Our pvtimeouts are ALL set to the default, as is the queue depth. We have found that with the timeout set to 60, 90... when there is a problem completing the I/O, the system waits, for as long as double the timeout value.
We have nothing but Open-9s in our XP256, connected in a SAN via switches.

Re: XP256: PV timeout and SCSI queue depth

Dave Wherry — Wed, 07 Feb 2001 23:31:19 GMT

We are all OPEN-8 and are running the default values. We've had no problems.

Re: XP256: PV timeout and SCSI queue depth

Philippe GAUQUELIN — Fri, 09 Feb 2001 08:14:10 GMT

We had also this problem with similar config (hub and not switch for SAN). I modify pvtimeout to 60 and qdepth to 2. Since that we have no more message. This value MUST be reset at each boot. So we add and /sbin/init.d/initdsk with /etc/rc.config.3/Sxxxinitdsk link.
Thind these script as attachment . Qdepth is configurable so that total qdepth is less than 256 on one FC link.
Here is the config file of attached script:
# config init de physical volumes pour XP256
# racine de l'adresse des disques
export ADDISK=c6t
# time out PV
export PVTO=60
# Q DEPTH "normale"
export QDEPTH=2
# Q DEPTH "speciale"
export SQDEPTH=2
# liste des device avec une QDEPTH speciale
# attention le nb total de qdepth doit etre < 256 par device
# pour temp tape4 tape2
#export LDEV="c13t3d3 c13t0d2"
export LDEV=""

Re: XP256: PV timeout and SCSI queue depth

Erik Tong — Wed, 20 Jun 2001 00:10:15 GMT

(Better late than never, I guess)

The recommended queue depths on the XP256 were limited by the controller boards on the XP, not the lun size or emulation type.

I can't remember the exact limit (they kept increasing it with newer versions of microcode), but the queue depth of 2 was based on 128 LUNs assigned per port and 2 FC ports per controller. This would indicate that 512 (128 * 2 * 2) was the maximum concurrent outstanding SCSI requests possible at any instant in time. If you did not use all 128 luns, or some luns had far less activity than others, you could shuffle the queue depths around (8 on one lun, 1 on seven others), just as long as you stayed under the overall limit. This MAY improve performance on some configurations. If you exceed the limit, (i believe) SCSI requests will be dropped and the host will be left waiting for a time out (some hosts will retry after that, but the delay is significant).

The overall limit has probably gone up, but this should give some back ground on why the queue depths have been recommended as such.

Setting the timeout is a good idea. The XP array, and most arrays, responds slowly reletive to a normal disk. Needless PV Link switch overs will degrade your performance.

Re: XP256: PV timeout and SCSI queue depth

Vincent Fleming — Wed, 20 Jun 2001 13:29:31 GMT

The XP can have up to 1024 tags active on a port. They are indeed shared between the LUNs.

When this limit is exceeded on an XP array, the host will be notified with a Q_FULL message when it attempts the next I/O. This will cause the host to hold onto (and delay) the I/O until the Q_FULL condition clears.

Note that if you're using another array (such as EMC or HDS), a Q_FULL will cause HP-UX to STOP USING TAGS until rebooted. This *will* destroy your performance.

The longer timeout and suggested short queue depth is because the array reorders the queue dynamically, and may delay execution of an I/O for quite a while - an I/O can be pushed back in the queue, if the array thinks that it is beneficial to do so.

The short (2) queue depth keeps this behavior to a minimum. The long timeout prevents HP-UX from failing the I/O if this queue reordering does occur.

Re: XP256: PV timeout and SCSI queue depth

ed skolnik — Fri, 02 Apr 2004 14:34:31 GMT

Does a any know how this applies to an HP EVA 5000 with HPUX 11i and that dreaded OS Windows 2000 or 2003?