Thanks a lot for your answer. Just to clarify, the reason I wanted to know this isn't really data validation. It's important for understanding of the results of my benchmarks. Since the program that I'm testing is a kind of iSCSI server, which needs to, for example, acknowledge writes based on user's policy pertaining as to how far the data had propagated, a read from a block with a write pending might be served from the server's cache, rather than going the full circle to the physical storage and back. Thus, understanding the percentage of such overlaps would be very important for me. So, I was wondering if there is any way of controlling this behavior. Thanks! On Thu, May 2, 2019 at 6:50 PM Elliott, Robert (Servers) <elliott@xxxxxxx> wrote: > > > > > -----Original Message----- > > From: fio-owner@xxxxxxxxxxxxxxx [mailto:fio-owner@xxxxxxxxxxxxxxx] On Behalf Of Left Right > > Sent: Wednesday, May 1, 2019 10:06 PM > > To: fio@xxxxxxxxxxxxxxx > > Subject: How do reads and writes overlap? > > > > I was looking for this info, online, but with no luck so far. What I'm > > really interested to know is this: > > > > * Say, you have a mixed sequential workload 75% reads, 25% writes with > > queue depth greater than one, then: > > ** are all writes executed before all reads? > > No. > > > ** are there any reads from the blocks not yet written to? > > Yes. > > > ** are there any reads performed against writes that are not ack'ed yet? > > Yes. > > > ** are there any writes performed to the same blocks that have a > > pending read request? > > Yes. > > * Does any of the answers change if we modify the ratio of reads to > > writes, sequential to random? > > No. > > Remember that fio is a primarily a workload generator, not a data > validation tool. It doesn't care what data it gets back from reads > (comparing all the bytes would take too much time, interfering with > performance). You can run special verify threads, but those are > separate from the normal read workloads. > >
