2016-11-14 04:11 AM
It's probably an age old question that comes up often around here. I have read through a number of posts about IOPS, and I *Think I have made some progress with my understanding, but I still have some questions.
Firstly with regards to IOPs calculations. I know that you have various technologies that can greatly assist with IOPs, such as PAM's and SSD's to create flash caches, etc etc. In our setup however we have none of those so it should be a straight up Number of Disks*IOP per disk.
For example I have a Shelf of SATA disks
20Disks * 75IOPs per disk = 1500
I realise that if it was all read Ops only you can take that stat of 1500, but for write you have some IOP penalty depending on RAID, which in this case is RAID-DP (2). Is my understanding about IOPs calculation correct?
I have been measuring Total IOPs off the aggregate (which is really just our shelf) and I have been getting readings of 4200IOPs. The read and write latency are still fine though ranging from 0.2 to 10ms, the latter I believe is still OK. I thought that A) the System should not be capable of anything more than that 1500IOPs, then how can it achieve the 4200IOPs? B) What other resources on your unit would start sufferring if your IOPS is high. Would your latency not be impacted as your unit is trying to deal with a surplus of IOPs?
I have not had any complaints thus far so I it has not warranted any urgent feedback, but I would like to understand better to support better.
2016-11-27 06:11 PM
hopefully this would help:
1. when we talk about IOPS like your sample, 75IOPS, we need to know what the io size is. coz if we dont know that, we can understand how big the workload is.
2. along with 75IOPS, is that fully ramdom? is it just read? or just write? or mixed? coz, if it's ramdom, the latency must be worse than sequential.
so in general, if you draw the IOPS and latency into 2-axle coordinate with condition 4/8k fully ramdom write, y is latency, x is IOPS. the curve really looks like y=x^2 (just for ur imagination, it's defiitely not precise). when IOPS reaches one certain point, latency goes infinity.
but in real world, we do have sequential and mixed IOs. so we definilty can go beyond your number, coz ur number should be the worst case.