Simple question regarding IOPS

Question

Hello, Im new here, and I dont know much about storage, so my apologies in advance for the basic nature of these questions.

I have Equallogic PS6100, RAID 6 accelerated, 85%7500rpm, 15%SSD. We have about 10 volumes in the default pool, that are sized from the maximum 15tb.

When we discuss maximum IOPS (Dell, or any other manufacturer), do they quote max IOPS per device, or per volume?

The topology is this: Host running CentOS and a ISCSI initiator, with a converged network adapter (10g) <-> Dell N4064 10g switch <-> Dell equallogic PS6100 (10g controller cards).

we notice that when we do process jobs on the linux host, we only get about 200 IOPS, 100mb/sec transfer rate, and average latency of 100ms. Im guessing that latency is not good. This process jobs are 100% read.

I guess my concern is that in the experimental analysis, the MAX IOP rate is much higher, and so Im just wondering if this is due to things not being optimally setup? Maybe I just need to spread this data out over multiple volumes to get the expedited performance numbers?

Thanks much for any and all help.

dwilliam62 · Answer

Hello,   Also the SANHQ manual has a chapter on performance monitoring that can help explain the SANHQ charts as well.   Don

dwilliam62 · Answer

Hello,

IOs are measured per pool or by member.

The experimental analysis is based on a relatively small blocksize, of 8K I believe. If you are doing large sequential reads with jumbo frames the blcoksize will be much larger.

The I/O rate is measured from the request time until it's acknowledged. When you have larger blocksizes it takes longer to send that data so the response time (latency) is longer. However, the benefit is that you get greater MB/sec in those cases. Small I/O size provide better I/O rates but less MB/sec. It's a balancing act with all storage.

Also with Linux you can set the readahead value to increase MB /sec on reads.

Make sure that all the Linux best practices are set

http://downloads.dell.com/solutions/storage-solution-resources/(3199-CD-L)RHEL-PSseries-Configuration.pdf

Also are you using the CNA with DCB enabled? If you have dedicated iSCSI switches the need for DCB is diminished.

Generally speaking more volumes do better than single mega volumes. In part because each volume negotiates a queue depth. Typically values are 64 to 128. This means on very busy volumes the queue may fill and pause additional I/O until the queue drains down.

Regards,

Don

nazimoo · Answer

Maybe I'm a storage neckbeard, but it really annoys me when people interchange Gb/s and GB/s.

dwilliam62 · Answer

Hello,   Well technically, he just did 'g'  for 'gig'    Which most take as 10Gb interfaces.     Don

dwilliam62 · Answer

Hello,  FYI: The 10GbE model of the 61xx series is a 6110.  The 6100 is GbE only.   Regards, Don

EqualLogic

Simple question regarding IOPS

Was this post helpful?