On Tue, 17 Jun 2014 15:20:53 +0200 Sebastian Parschauer <sebastian.riemer@xxxxxxxxxxxxxxxx> wrote: > Hi Neil, > > have you already had the time to look at these patches? I have now looked through them at last. Sorry for the delay - there was a week of leave in there and various other distractions. The improved accounting is probably a good idea. The extra malloc for each request is a bit unpleasant. For raid1 and raid10 we could use the r1_bio they already allocate to store the extra data. For raid5/6 we could possibly do something similar though the results might be a little less precise. For raid0 and linear I'd rather not intervene at all. The underlying devices will report accurate statistics which can be combined to get correct statistics for the overall devices. The request function support is a bit of a mix. In some ways it is quite nice and localised. However I'm not thrilled with the extra work queue, and I don't really understand why md_process_request() allocates a new md_request_clone and clones all the bios... I wonder if that is only needed for RAID0, not for higher levels. I still don't think any of this should be applied to RAID0. I asked about that before and you didn't really provide an explanation. RAID0 simply maps the bio to a different device and offset and sends it down. So the queue management on the underlying device should be able to provide all the fairness etc that you need. So if you have RAID0 over RAID1, then I can understand the desire to use requestfn rather than make_requestfn on the RAID1, but not on the RAID0. Have you tried that configuration? i.e. the RAID0 made without --use-requestfn and the RAID1s made with it? Does that work well as well as with both arrays using requestfn?? The graphs you posted subsequently certainly look impressive, but I really want to make sure I understand what is really needed. Thanks, NeilBrown > > We have this in production with cfq scheduler already on two storage > servers and customers noticed a huge performance boost. Also our > monitoring shows higher and smoother iop numbers. > > Cheers, > Sebastian > > > On 04.06.2014 19:09, Sebastian Parschauer wrote: > > Patches apply to linux v3.15-rc8 and mdadm master commit 8a3544f895. > > > > Florian-Ewald Mueller (3): > > md: complete bio accounting and add io_latency extension > > md: introduce request function mode support > > md: handle IO latency accounting in rqfn mode > > > > Sebastian Parschauer (1): > > mdadm: introduce '--use-requestfn' create/assembly option > >
Attachment:
signature.asc
Description: PGP signature
