SUMMARY: sd_max_throttle settings in /etc/system and in jni configuration file

• Previous message: Cirus Gaspar: "SUMMARY: SunCluster 3.0 behaviour at shutdown"
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Date: Fri, 28 Nov 2003 09:36:35 -0000
To: "SUN Managers (E-mail)" <sunmanagers@sunmanagers.org>

All...

Apologies for the late delivery of this posting. I forgot I hadn't sent it
out. The original posting is shown at the end of this mail for reference.

Many thanks to the following who. There were others who replied who didn't
want to be named. Many thanks to you all:

Johan Hartzenberg [mailto:jhartzen@csc.com]
Kris Briscoe [krisb@tti-telecom.com]
Gunnar Brasack [gbr@asv.de]

There was some discussion as to whether the sd_max_throttle setting was EMC
only. We are using XP512 (rebadged HDS arrays) so this may not be a relevant
parameter for us.

Others reported that that had had discussion with sun about the
sd_max_throttle parameter in the /etc/system file. The advice was <256 /
number of LUN's>.

There was some discussion around disabling DMP to see if performance
declined/improved etc. If it improved, maybe investigate adding additional
paths.

Basically - we haven't resolved the problem and investigations continue. We
tested values of 2 (/etc/system) & 64 (JNI file), 8&8, 4&4, 2&2, 20&20. The
throughput never changed more than a few percent, but always in the downward
direction. We have therefore concluded that this is not a significant cause
of the problems and set the values back to their defaults of 2&64.

Miraculously, our slow disk service times have now gone away but nobody knows
why. Always worrying! The SAN team made a few tweaks to the infrastructure
but wouldn't admit to changing anything significant.

Also, HP have made some changes on the array to cater for other projects but
say that this would not affect us.

The box itself is now under low stress but the throughput has not increased.
There are some Oracle tweaks our DBA is going to test out and we are going to
investigate the network outside of the host itself.

If I come up with anything devastating, I'll publish a further summary.

thanks and best regards

Tony Miller

======================Original
Posting========================================
All...

I have a Solaris-8 F15K domain with 2 x JNI FCX2-6562 dual port adapters
and V5.3 of the JNI driver.

The JNI's are connected to brocade-3900 switches and are forced to 2Gb. An
XP512 disk array is also connected to this dedicated disk SAN. We have 120
LUNS (Open'E's) presented to each port of the first card (i.e.., 240 LUN's
in total). VxVm DMP is used and we can see these 2 x 120 LUNS down the
alternate paths on the second card.

The LUNs are under VxVm control, (V3.5) with VxFs V3.5 file systems on top.
We have done some file system tuning via the /etc/vx/tunefstab and set the
relevant mount points up with read_pref_io=64k,read_nstream=1.

I wont go into much detail but we are experiencing some disk write
throughput issues. This is under investigation but specifically the issue
relates to a dramatic throughput drop off in a highly parallel workload
environment.

Shown below is the tail end of the JNI configuration file - for the last
two LUNs only:

jnic146x3-target0_lun118_hba="jnic146x3";
jnic146x3-target0_lun118_throttle=64;
jnic146x3-target0_lun119_hba="jnic146x3";
jnic146x3-target0_lun119_throttle=64;

Below is an extract of my /etc/system file:

* Allow SCSI transfers up to 8MB on VXVM layers
set vxio: vol_maxio=16384
      : :
      : :
*
* Required for XP512
*
set sd:sd_io_time=0x3c
set sd:sd_max_throttle=2
set maxphys=8388608

I have searched the archives and various other mailing lists. Several
articles refer to the /etc/system file entries as being the 'proffered'
ones with 'optimal performance'.

I have a specific question relating to the lun throttle settings (of 64 in
the JNI driver file) and its interaction with the sd_max_throttle setting
of 2 in the /etc/system file. I checked the Solaris-8 kernel tunables guide
but
couldn't find anything useful with regard to this.

Question 1 - do the LUN level throttle settings (in the JNI config file)
override those in the /etc/system config file or vice versa? Or are they
independent (i.e.., the /etc/system file entries relating to the direct attach
SCSI disks only).

Question 2 - Are these suitable throttle settings (for 120 luns per HBA
port). It has been suggested that values of 4-6 would be more appropriate
(1024
for an XP512 divided by 120 LUN's per port = 8.5)

Sorry to pose such vague questions, but your advice is appreciated.

Many thanaks and best regards

Tony Miller

===================End Of Original
Posting========================================

+----------------------------------------+
| TONY MILLER
| Team Leader : Technical Projects,
| VODAFONE LTD,
| Derby House,
| Newbury Business Park,
| Newbury, Berkshire.
|
| Phone +44 (0)1635-677687(local)
| Mobile +44 (0)7766-028752
| Email anthony.miller@vf.vodafone.co.uk
| FAX +44 (0)1635-233517
+-------------------------------------------
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers

• Previous message: Cirus Gaspar: "SUMMARY: SunCluster 3.0 behaviour at shutdown"
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]