proper crescent
As I know of, NetApp volume block size is 4k, so, the IOPS refers to how many 4k blocks I/O's can be completed per second. Oracle uses 8K as the block size, so, if we see 32 IOPS in AIQUM on a volume used by Oracle data file , for instance, that would be equivalent to 16 IOPS in Oracle.
What will this difference cause any impacts on our performance analysis if any?
formal reef
I'm curious about this as well.
Somewhat related—our old Nabox (version 2.x) had a QoS dashboard that gave us a rough estimate of IOPS from the client’s perspective. I found that really useful, but it seems to be missing in the latest version. Unfortunately, my Prometheus query skills aren’t quite up to the task of recreating it. 😅
green perch
IO size is independent of block size. If you do 30 read requests of 1 mb each, you get 30 IOs from the client to the storage. WAFL will then have to "convert" this to 7500 4k IOs, but then at the disk layer it will get converted down to much less again (becauseWAFL coalesces writes to the same disk, i.e. instead of writing 100 4k blocks to the disk it will write one 400k range all at once).
I am not 100% sure which number AIQUM shows but I guess it's probably the first one, i.e. the number of client I/Os coming into the system
proper crescent
I believe AIQUM's IOPS is based on the Netapp block size 4k.
One of reasons I bring it up is because we balance workloads based on IOPS # showed on nodes in the cluster. If too many IOPS showed on a node comparing to the other nodes, we tend to move the volume to other nodes.
I am kind of thinking we should balace based on latencys.
Will you make this kind of balancing regularly based on IOPS even with no performance issues found?
green perch
that's a pretty secure password 😉
echo nebula
Op size is just what ONTAP reports, same as using the workload object in statistics or qos stats commands.