On Tue, Feb 21, 2023 at 04:51:21PM -0700, Jens Axboe wrote: > On 2/21/23 2:59?PM, Luis Chamberlain wrote: > > On Fri, Feb 17, 2023 at 08:22:15PM +0530, Pankaj Raghav wrote: > >> I will park this effort as blk-mq doesn't improve the performance for brd, > >> and we can retain the submit_bio interface. > > > > I am not sure if the feedback was intended to suggest we shouldn't do > > the blk-mq conversion, but rather explain why in some workloads it > > may not be as good as the old submit_bio() interface. Probably low > > hanging fruit, if we *really* wanted to provide parity for the odd > > workloads. > > > > If we *mostly* we see better performance with blk-mq it would seem > > likely reasonable to merge. Dozens of drivers were converted to blk-mq > > and *most* without *any* performance justification on them. I think > > ming's was the commit log that had the most elaborate performacne > > metrics and I think it also showed some *minor* slowdown on some > > workloads, but the dramatic gains made it worthwhile. > > > > Most of the conversions to blk-mq didn't even have *any* metrics posted. > > You're comparing apples and oranges. I don't want to get into (fairly) > ancient history at this point, but the original implementation was honed > with the nvme conversion - which is the most performant driver/hardware > we have available. > > Converting something that doesn't need a scheduler, doesn't need > timeouts, doesn't benefit from merging, doesn't need tagging etc doesn't > make a lot of sense. If you need none of that, *of course* you're going > to see a slowdown from doing all of these extra things by default. > That's pretty obvious. > > This isn't about workloads at all. I'm not arguing mq design over-architecture for simple devices. It is a given that if one doesn't need some of those things surely they can create a minor delta loss in performance. I'm asking and suggesting that despite a few workloads being affected with a *minor delta* for brd for mq conversion if the huge gains possible on some *other* workloads suffice for it to be converted over. We're talking about + ~125% performance boost benefit for randreads. Luis
