From the iostats documentation, you can see that ms_read + ms_write is “the total number of milliseconds spent by all reads (as measured from __make_request() to end_that_request_last()).” This doesn’t get incremented until the read completes. However, ms_weighted increases as the reads are in progress. Various scenarios can lead to the numbers being bigger or smaller. Remember also that the numbers in my sample are all made relative to the previous line.

Milliseconds spent reading and writing are indeed the sum of all requests, so they are higher than wall-clock time if there is parallelism. And you can apply Little’s Law to these numbers to compute the parallelism and a bunch of other stuff. I think I have a draft blog post (part 2 of my previous post you referenced) on that.

In your example here, what does it mean when (ms_read + ms_write) > ms_weighted, and what does the opposite mean? I see cases of both — line 3, ms_weighted is almost 2x higher; third line from bottom, it is about 15% lower.

Also, would I be correct to say that ms_reading|writing are sum of per-thread stats (similar to ms_weighted), such that they are often higher than elapsed wall-clock time? If so, is the relationship between ms_doing_io (wall-clock time) and ms_weighted (per-thread) approximately indicative of the amount of IO parallelization happening?