Disk Performance issue

[Shane Russell]

Hi,

I am experiencing disk preformance issues on our CRM system. Disk utilization is often running above 90% according to glance & sar. Performance Manager narrows it down to 3 disks (LUNs) in one Volume Group & one Logical Volume.

The Logical Volume is a 3 way stripe on 3 x 70G LUNs. The 70G LUNs are striped 8-way on a DMX 3000. So it is stripe on stripe.

Should I increase the LV to a 4-way stripe? or 6-way or 8-way?

Can LVM relayout an LV the way VxVM can?

Any other pearls of wisdom.

Further Information:
* Both paths are experiencing heavy utilization. The HBA's, FC cable & switch ports are all 2G.

* The HBA ports in the FC switches are only 5% used. The FA ports in the FC switches are 35% used.

regards,
Shane

[Steven E. Protter · Accepted solution]

Shalom Shane,

Dis utilization is not a good measure to judge whether to take action.

Application response is a good issue. If the application responds properly and everyone is happy, find something else to do.

In general there is one common problem in which raid 5 and other striping methodology are used to host write intensive databases. Raid 5 inherently writes more slowly because the data needs to be written in many places. Moving write intensive applications to Raid 1 or Raid 1/0 storage can dramatically improve write performance.

If you must remain in the striped world having the most disks is the accepted norm for good performance. Any of the changes recommended above are usually handled on the disk array by LUN reconfiguration.

SEP

[Shane Russell]

Thanks SEP,

I am not using RAID5. It is just a 3-way stripe. The 70G LUNs are mirrored in the DMX.

The 70G LUNs are striped 8-way in the DMX, so ideally I'd like to relayout at the OS layer.

I could create a new 4-way LV on 4 new 70G LUNs and move the data to it - what do you think?

regards,
Shane

[Zinky]

Shane,

Unfortunately you cannot do an online relayout or better yet a "plex replacement" as can be done with VxVM.

As far as performance, here are my suggestions based on our experiences:

1. Use 4 or 8 way stripes.. (for CRM use a stripe width of 64). With Hitachi based frames and EMC, we never noticed any difference between RAID5 and RAID10 LUN components on these stripes..

2. Make sure your filesystems are mounted with DirectIO enabled (fstab ... convosync=direct...)

3. And make sure your buffer cache is no more than 800 to 1600 MB (IF your server purely does DB and APP serving...)

[Shane Russell]

Thanks all