topic Re: Any idea about improve I/O performance? in Operating System - HP-UX

Any idea about improve I/O performance?

Lukas Grijander — Tue, 19 Jun 2001 16:07:36 GMT

I've an ORACLE db across 4 fc60 LUNS (data, idx, rbs, etc balanced across all the LUNS).

Each LUN has 4 disks on RAID-5, but I'm planning to change to RAID-0/1 (as soon as posible).

Any idea about fc60 stripe size of LUNS? Now is 8 Kb.

Any idea about filesystem block size? Now is 8 Kb.

Thanks in advance.

Re: Any idea about improve I/O performance?

Edgar Matzinger — Wed, 20 Jun 2001 07:59:53 GMT

Hi Rafael,

according the documentation is "The optimum stripe segment size is the smallest size that will rarely force I/Os to a second stripe." If you're using 4 disks to create RAID 0/1 striping, i.e. 2 disks for the original data and 2 disks to hold the mirror. And if all I/Os are 8KByte, then a stripe size of 4KByte would be the optimum. But if the vxfs driver would group multiple 8KB I/Os into 1 64KB I/O, then the optimum would be 32KB. Because all stripes needed for the I/O request can be read (or written) simultaneously. Please note that the largest blocksize vxfs supports is 8KB.

HTH, cu l8r, Edgar.

Re: Any idea about improve I/O performance?

Alexander M. Ermes — Wed, 20 Jun 2001 08:42:31 GMT

Hi there.
You talk about LUN's. If you use several controler interfaces, psread it thru the different interfaces as well. Otherwise your your controler will be the bottleneck.
Rgds
Alexander M. Ermes

Re: Any idea about improve I/O performance?

Lukas Grijander — Wed, 20 Jun 2001 10:39:09 GMT

Hi Alexander.

I use 2 controlers.

thanks.

Re: Any idea about improve I/O performance?

Lukas Grijander — Wed, 20 Jun 2001 10:49:27 GMT

Hi Edgar.

Should I understand that ...?

1.- The fc60 stripe_size must be block_size / (number of data disks).

2.- Vxfs has an 8 kb bs. I think JFS On_line bs may be larger than 8 kb. I read somewhere "vxfs (with JFS On_line) likes bs of 64 kb".

In any case I haven't JFS On_line.

Thanks.

Re: Any idea about improve I/O performance?

Bill McNAMARA_1 — Thu, 21 Jun 2001 09:43:20 GMT

Use two controllers to access the luns.
ie don't put all data io traffic down just one FC60 controller.
When creating the LUN, make sure that you create it across enclosures (or split busses)

Split bus mode will be faster, only 5 disks/bus.
Full bus mode will have 10 disk/bus.

What does amdsp -a show

Is the FC60 the only thing on the fc loop?

Later,
Bill

Re: Any idea about improve I/O performance?

MARTINACHE — Thu, 21 Jun 2001 13:24:07 GMT

Hi,

If you want to improve IO perf, you should purchase Online JFS.
With this soft, you mount oracle filesystems with "-mincache=direct" option.
This will give perf such as raw-device datafiles.

Regards,

Patrice.

Re: Any idea about improve I/O performance?

Lukas Grijander — Thu, 21 Jun 2001 15:16:15 GMT

Thanks Patrice.

I don't know if I have understood so well, sorry, but I'm going to try giving you more information.

There are 5 HP9000 "against" 1 fc60.

In the fc60 there are 52 disks, 16 of them are used by the machine of the question (ORACLE, etc).

These 16 are accross the 6 scsi channels.
scsi 1 - 2 disks
scsi 2 - 2 disks
scsi 3 - 4 disks
scsi 4 - 4 disks
scsi 5 - 2 disks
scsi 6 - 2 disks

Every LUN was created across different SCSI.

Is enought or you need more information?

Thanks.

Re: Any idea about improve I/O performance?

Phil Miesle — Thu, 28 Jun 2001 19:43:31 GMT

Hi there,

Oracle has a methodology called "SAME" (Stripe and Mirror Everything). You may be able to find the whitepaper on this (it was presented at Open World last November); if not let me know and I'll post a copy.

Basically, you should create a single LUN which is striped and mirrored; in your case, it will be 8 disks plus 8 disks. Keeping the controllers balanced it will be the disks on channels 1-3 mirrored to the disks on channels 4-6.

You asked about stripe depth. The SAME methodolgy suggests as large a stripe depth as possible (up to 1 MB) in order to optimise multiblock operations.

I will suggest that if you archive you should create two LUNs, one for archives and one for the rest of the database. It would be sufficient to allocate 2 of the 16 disks to your archives (I'd mirror them), this depends of course on how big your archive area needs to be compared with your disk size.

This configuration should allow for good performance (90-95% of optimal in most situations). In order to squeeze the last 5-10% out of the system, you will need to I/O profile your application, but you'll be limited only having 16 disks.

You can make things a lot more complicated and try to fine-tune the performance, but unfortunately as the nature of your application usage changes over time it is likely your disk configuration would need to change as well.

Using the SAME methodology eliminates a lot of the administrative nonsense that you're trying to go through to achieve good performance.

hth,
:-Phil

Re: Any idea about improve I/O performance?

Stefan Farrelly — Thu, 28 Jun 2001 20:36:18 GMT

Weve recently being doing the same tests on an FC60 with 3 SC10's for an Oracle DB.

After much testing the fastest throughput we got was; 166 MB/sec

This was using RAID1 on the FC60 (not RAID0/1) with 4k stripe size, and using lvm striping with a 64k blocksize. Both stripe sizes are the optimum. The FC60 manual quotes 170MB/s as the max throughput so were getting very close!

Once you setup your lvols this way test them with;

time dd if=/dev/vgXX/rlvolYY of=/dev/null bs=1024k count=50

and see if you can get the time down to 0.3 sec (50MB/0.3=166 MB/s). For some as yet unknown reason using RAID0/1 was a bit slower (just over 120MB/s).

Re: Any idea about improve I/O performance?

Lukas Grijander — Fri, 29 Jun 2001 06:21:48 GMT

Hi Phil.

First, thank you.

Second, I was looking for the SAME methodology at ORACLE site but I couldn?t find anything (what a "same" ...). Please send me the whitepaper if you can.

Thanks.

Re: Any idea about improve I/O performance?

Phil Miesle — Fri, 29 Jun 2001 08:08:22 GMT

Rafael (et. al.),

Attached is the SAME methodology paper presented at OpenWorld. I hope you find it useful!

:-Phil