Thanks to the following people who replied to my problem:-
Paul O'Sullivan
Andy Williams
Rodrigo Poblete
alan_at_nabeth.cxo.dec.com
Viktor Holmberg
These replies are detailed below. I'm still not getting Ultra SCSI
performance through these controllers and have tried multiple of tests. I
have escalated to COMPAQ but so far they can not understand why I'm only
getting 8-10M per second on reads and 14MB for writes. They are still
investigating.
Thanks
Elaine
From: Paul_O'SULLIVAN_at_paribas.com [mailto:Paul_O'SULLIVAN_at_paribas.com]
Sent: Tuesday, May 11, 1999 5:56 PM
To: Elaine.Andrews_at_yes.optus.com.au
Subject: Re: Ultra SCSI
Elaine,
What is the cache size in the controllers? Is it read/write or read?
Are the reads sequential?
From: Andy_WILLIAMS_at_paribas.com [mailto:Andy_WILLIAMS_at_paribas.com]
Sent: Tuesday, May 11, 1999 8:49 PM
To: Elaine.Andrews_at_yes.optus.com.au
Subject: Performance stuff..
I wouldn't use cat for benchmarking. Why not use dd to /dev/null ?
At least you'll have a comparable benchmark.
If you still get poor read performance; have a play with the following
On the HSZ controller; run vtdpy.
This gives you details information on cache hits and transfer sizes etc
>From the main screen type CTRL/C
You can then use display cache, default,device, status
This may give some clues as to whats happening. Look for low cache hits etc.
I think COMPAQ have a tool which allows you to collect all this data and
stick
it
in a spreadsheet.
I guess you have the stripes going across the bus ?
Also try setting a preffered read plex.
From: Rodrigo Poblete [mailto:rpoblete_at_gmd.com.pe]
Sent: Tuesday, May 11, 1999 11:19 PM
To: 'Elaine Andrews'
Subject: RE: Ultra SCSI
Hi elaine:
If you want to set the KZPBA-CB to work at USCSI speed
you have to set the board to that speed runnning the eeromcfg configuration
utility that comes in the Alpha Firmware upgrade CD.
If you don't set it (oor at least verify it), most probably your board is
operating at differential speed only.
The documentation on how to load this utility comes in the CD in the DOC
directory in the 8200_8400 firmware upgrade procedure manual
Good luck
Rodrigo Poblete
From: alan_at_nabeth.cxo.dec.com [mailto:alan_at_nabeth.cxo.dec.com]
Sent: Wednesday, May 12, 1999 12:52 AM
To: Elaine.Andrews_at_yes.optus.com.au
Subject: Re: Ultra SCSI
Is the "cat" test through a file system or raw disk? If raw disk,
why not use dd like you did for the write test? Cat(1) uses what-
ever I/O size it uses. dd(1) lets you chose the I/O size. That
may be part of the difference. The limitation that any raw I/O
test has over a file system test is that file systems can do
read-ahead. This can allow for higher data rates than you'd get
on the purely serial reads to a raw device.
It is worth noting that on a six member stripe set of 9 GB disks
on an HSZ70 I can get ~14 MB/sec reading through AdvFS. When I
stripe the files across three such controller stripe sets using
AdvFS striping, I get closer to 16 MB/sec. In my case all three
stripe sets are on the same array controller.
From: Viktor Holmberg [mailto:Viktor.Holmberg_at_uk.abnamro.com]
Sent: Thursday, May 13, 1999 12:00 AM
To: Elaine.Andrews
Subject: Re: Ultra SCSI
Hi Elaine,
Whar are you using to measure the throughput? monitor ?
OK the maximum speed of the UWD scsi is about 30M per second after the
scsi negotation is taken into consideration, not 40M per sec.
How did you do the dd test? Did you read a file on one of the volumes and
write it to another? If you did this you would have created contention on
the scsi buses.
A cat of a small file isn't very good for performance measurements size
you will have the initial delay of starting the process of the cat,
reading the inode information, and most importantly the delay of the
screen you are catting to (keaterm, reflections, xterm etc)
If you do
dd if=/file_system/big_file of=/dev/null bs=Xk
where X is a variable block size. You will get a very good idea of the
performance. Changing ther block size will allow you to tune a little as
well. The UFS and AdvFS have an atomic block size of 8k, however AdvFS
will try and consolidate the reads and writes into bigger more efficent
blocks.
Better still you can
dd if=/dev/vol/diskgroup/volume of=/dev/null bs=Xk
This will give you the raw performace of the disks without the filesystem
overheads. Since you are reading you don't need to unmount the volumes.
The of course you can dd the individual disks.
Good luck,
Viktor
This is the original message I sent!
> -----Original Message-----
> From: Elaine Andrews
> Sent: Tuesday, May 11, 1999 5:10 PM
> To: 'tru64-unix-managers_at_ornl.gov'
> Subject: Ultra SCSI
>
> Hello, this is my first e-mail to you guys and I was wondering if anybody
> can help.
>
> I've just set up an Alpha 8400 with 4 HSZ70's connecting through 4
> Ultra-Wide differential SCSI controllers (KZPBA-CB) under Digital Unix
> 4.0D. To test the performance of these super fast disks I set up a three
> disk stripeset (using UltraWide disks) mirrored with another three disk
> stripeset using LSM over 2 different controllers. The writeback cache and
> the transfer size has been set.
>
> When doing some benchmarks it appears they are not performing as they
> should!
>
> Using dd for writes they are reaching about 14MB per second which seems
> okay. However, when using cat for reads (pre cache) I'm only getting about
> 8MB per volume which works out to be about 1.3MB per physical disk. This
> seems extremely slow. Has anybody ever come across this before. Am I
> missing something?
>
> Thanks
> Elaine Andrews
>
Received on Wed Jun 02 1999 - 04:51:44 NZST