Disk IOPS - Navi Analyzer

Navi Analyzer counts utilization of a resource as busy time / (busy time + idle time). Busy is defined as when one or more requests are active or queued at the resource. So 40% utilization means the LUN has 1 or more active requests 40% of the time. If a burst of requests arrive "all at once", they may be queued briefly at the LUN level, then distributed to the drives that support that LUN, so each drive gets several requests and each drive starts to work on the first one in its queue. So the "LUN" at that moment is very busy working on several requests in parallel, and also has several other requests queued up. Once that burst of requests is completed, the LUN is counted as idle, in your case, 60% of the time. Key point: utilization is NOT showing the percent of maximum performance capability.

As previously mentioned, drive IOPS rules of thumb like "150" are only guidelines, not maximums or limits. The intention is to give you a single number to use for conservative estimates of how many drives are needed for your workload, with preference for best response time. As such, that number must cover all cases. It is generally based on the capability of the drive when presented with a low thread count small block random workload spread across at least 1/2 of the drive GB capacity. This IO profile requires the drive to do significant seeks, and seeks take time.

If your workload is more sequential, or highly localized, there may be little or no seeking. Higher thread counts at the drive allow the drive to reorder the requests to minimize seek distance, which minimizes service time and maximizes throughput. With a pure sequential small block workload, with a few threads so the drive always has the next request in queue as it executes the current request, it can read those small blocks as fast as the platter turns, no missed spins, no seeks. With a workload like that the drive can deliver each piece of data in 1 ms without too much strain, and 1 ms service time means 1000 IOPS.

A pure load like this is rare and this performance should not be expected, but it can occur briefly during the natural bursts of a normal workload. That is why you may see disk IOPS far higher than the rules of thumb, even though utilization is much less than 100%