Re: troubleshooting slow I/O performance.

From: Mladen Gogala <>
Date: Tue, 8 May 2018 23:05:39 -0400
Message-ID: <>

If 4TB SATA drives are used for a high end Linux cluster, and 5 nodes certainly sounds like the high end configuration, then the configuration is ill-conceived from the very start. Add some RAID level, which always adds some overhead, and you get 10ms. Now, the question is what does that RAC need to do? If that is to be a highly available OLTP configuration, that is probably fine. If, on the other hand, this was supposed a data warehouse, and 5 node RAC certainly sounds like a DW configuration, then there is a problem.

Fortunately, ASM allows easy migration. You simply add a fast LUN, on XtremIO, Pure or something similar, wait for re-balance to complete and then drop one of the slow LUN's.  There are only 3 questions:

  1. How fast IO do you need? How quickly do you need to perform a full table scan of a 20GB table without parallelism?
  2. How much money can you spend? The old wisdom says that disks are cheap.
  3. Do you feel lucky?

I would do bonnie++ or iozone on the disks, see whether the current configurations gives you desired result and change configuration to something better, if the answer to the previous question is negative.

Go ahead, make my data!

On 05/08/2018 03:17 AM, Stefan Knecht wrote:
> I haven't spend that much time looking around, but it seems to paint
> the picture that most of those spin at 7200 RPM. Which leads to a seek
> time of approx 10ms. If you're hitting the array hard enough so that
> you're bypassing any caches and other smart gimmicks (which slob is
> certainly doing) - you would be essentially limited by the seek time
> of the spindles themselves. So this sounds about right to me.

Mladen Gogala
Database Consultant
Tel: (347) 321-1217

Received on Wed May 09 2018 - 05:05:39 CEST

Original text of this message