LVM Striping of LUSE's?

[Dan Hoyt]

Would I see a benefit if I did LVM striping across LUSE's?

I just created 10 OPEN-E*3 LUSE's (Made up of expanded dispersed OPEN-E's spread across 3 parity groups), and thought I might get better performance if I stripe them at the LVM.

XP512

Any input conserning question is welcome.

Thanks,
Dan

[Vincent Fleming]

You will generally see better performance in HPUX with LVM striping, as the OS "sees" more than one device, and can queue I/Os to the individual devices.

For example, if you have one device and no striping, you can queue only 8 (If I remember correctly) I/Os at a time to that device; however, if you stripe over 10 devices, you can queue 80 I/Os (8 per device).

The effect of this is significantly better performance and higher throughput on heavily loaded systems. (lighter I/O loads don't really take advantage of the queueing - you have to fill the queues!)

Now, you have to consider the geometry of what you've created. Think of a senario where you have 10 LUNs on the same array group (or drive, if it helps to think of it that way) - if you stripe them, it means that I'Os will be evenly spread over the LUNs, which means that they'll be evenly spread over the "drive".

This would be bad.

Why? Head movement. By doing such, you would be seeking the heads on the drive for every I/O... long seeks. That's slow.

So, you would most likely be better off making one LUSE on each array group (ie: no disperse), and stripe those 3 LUNs with LVM.

One last thing - LUSE on the XP512 is not the most efficient thing - your HPUX host usually has more CPU power available for such operations, so unless it's really heavily loaded, you will generally get better performance using regular LUNs and having the host stripe them.

I hope this helps. Good luck,

Vince

[Dan Hoyt]

Thanks for your help!