Subscribe

High IOPS values with SQLIO

Hello,

We are trying to make things clear about performance, so we made set of tests with SQLIO which were running on single FAS2240 with 28 FC disks in single aggregate (raid group size is max). We connected FC LUN directly to server (8gbit), without FC switches (to eliminate bottlenecks on other devices). We know basics, that single disk can't handle more that 130 IOPS (with normal latency), so results were quite strange:

We tested workload of 4k blocks, 8 threads, random write and got result around 20000 IOPS (from SQLSIO result and also from sysstat -x output). Obviously that is not the most logical result, since system produced around 800 IOPS per disk in aggregate. Now we are trying to figure out from where came this difference between theoretical and actual measured IOPS. Can someone please help me understand those results?

Thanks!

Regards,

Rozle

Re: High IOPS values with SQLIO

sysstat and client-side monitoring will give you the number of protocol level operations.  These are not the same as disk ops.  ONTAP will build up the writes in memory (and NVRAM) and commit writes when the NVRAM buffers are full.  these will be organised in memory to form contiguous chunks of data, and written in chunks.

Try using statit to get more info around the number of IOPS each disk is doing - that should make more sense.

Re: High IOPS values with SQLIO

Hi, Can you please share your sqlio commands?

Thanks

-Vishal


rozle_palcar wrote:

Hello,

 

We are trying to make things clear about performance, so we made set of tests with SQLIO which were running on single FAS2240 with 28 FC disks in single aggregate (raid group size is max). We connected FC LUN directly to server (8gbit), without FC switches (to eliminate bottlenecks on other devices). We know basics, that single disk can't handle more that 130 IOPS (with normal latency), so results were quite strange:

 

We tested workload of 4k blocks, 8 threads, random write and got result around 20000 IOPS (from SQLSIO result and also from sysstat -x output). Obviously that is not the most logical result, since system produced around 800 IOPS per disk in aggregate. Now we are trying to figure out from where came this difference between theoretical and actual measured IOPS. Can someone please help me understand those results?

 

Thanks!

 

Regards,

Rozle