extent based striping gain performance?

[Lynn Hsu_2]

Hi,
We are running data base app on L class with 2 FC10s. disks are 10-10 mirrored. The LVM configuration is PVG-strict/distributed. The PE size is default at 4MB. Can anyone point out what this distributed (extent based striping) do to performance? The FC has 1 controller . Will a contiguos allocation be OK in this case? Thanks.


Lynn

[Steven E. Protter]

I wonder, do you have a performance problem. Perhaps measure it with sar -d or http://www.hpux.ws/system.perf.sh scripts.

Your best bet to improve performance is to handle striping exclusively on the disk array and free HP-UX from these worries.

You might to get a better answer want to let us know what kind of transactions happen on the disk.

if its an oracle database with a lot of writes, striping is a bad idea and will slow down writes. If the data is static or mostly read data then striping will not have such an adverse impact on performance.

SEP

[A. Clay Stephenson]

You are going to find that extent-based striping does very little to enhance performance. The problem is that the smallest possible PE size (1MB) is still much too large to efficiently distribute the i/o. Moreover, 1MB PE's severely limit the maximum size of a PV. Sadly, a good stripe size (64,128, and 256KB at an absolute maximum) is just not possible with extent-based striping.

[Lynn Hsu_2]

We use "progress" DB. Our app is to process credit card athorization then handle the transaction. We are facing performance issue as well as disk space issue. So an upgrade plan is on the way. The new disks are DS2400 arrays.

Lynn

[Mark Ellzey]

Lynn,

Although quite a few years ago, I set up a Progress DB on an HP array using RAID 0/1. I configured the array to have the database on one set of spindles and the .bi files on another, each one mirrored. We weren't using .ai files. I also had 2 RAID controllers, so count route the DB and .bi writes through different controllers.

This supported about 200 users on an old G series box, with no performance issues. Even month-end processing was pretty quick.

This may give you some ideas.

Regards,
Mark

[Ted Buis]

For those not familiar with the FC10, it is not an array, but just a FC JBOD enclosure that holds 10 disks. The best answer to your question is, "It depends." Any performance gains could depend greatly on the application. Do you have just one FC Host Bus Adapter (HBA) for both enclosures, or one for each? I would recommend one FC HBA for each enclosure. This would eliminate a single point of failure and help performance somewhat.

The next step would be to better understand where the pressure is on the I/O. Is it just on one physical disk? Is the I/O pressure from page-outs? Would more buffer cache help? Is there enough CPU to support more I/Os or more overhead? What are the mount options on your filesystems? It is possible that the full VxVM could provide a better solution than LVM, but there is no way to know without much more information?

[Zinky]

If you're using the DS2403 Fibre JBOD - then I will strongly recomend you use Full VxVM (or Veritas Foundation Suite for 11.11 - If you're running 11.23 it will be Storage Foundation). If you already have OnlineJFS (aka VxFS) - that is the other half of Foundation Suite and you'll only have to pay for Full VxVM.

With VxVM you have more tunables to create the perfect mirrored stripe (or striped mirrors) that best suites you. You'll also gain a number of backup/recovery/protection techniques not available with LVM.

[Ted Buis]

The FC10 is not the same as the ds2405 (Nelson made a typo), but the operate similarly. The FC10 was a 10 disk enclosure and the ds2405 was a 15 disk enclosure. However, I think Nelson's comments are totally appropriate for the FC10 too.

[Lynn Hsu_2]

enough info. Thanks!