Jens Axboe@axboe·Apr 1I still dream of improving IO scheduling like this. https://lore.kernel.org/lkml/e9c15e2066177c3efdfe6d134cf9c80b5e8f8d1b.1491076459.git.osandov@fb.com/… CC @omsandov6859
Mark Callaghan@MarkCallaghanDB·Apr 1Replying to @axboe and @omsandov"When applied to network routers, RED probabilistically either marks packets with ECN or drops them," How does that map to disk IO?11
Jens Axboe@axboe·Apr 1Replying to @MarkCallaghanDB and @omsandovPretty simple - if you run out of requests trying to read or write data, then the IO is simply dropped rather than queued. That prevents congestion on the device.17
Mark Callaghan@MarkCallaghanDB·Apr 1Replying to @axboe and @omsandovFor buffered IO writes would the error be return on the write by userland or to the kernel on writeback?11
Jens Axboe@axboe·Apr 1Replying to @MarkCallaghanDB and @omsandovJust EIO it and let the application read or write it somewhere else.22
Peter Geoghegan@petervgeogheganReplying to @axboe @MarkCallaghanDB and @omsandovThe fsync philosophy5:54 PM · Apr 2, 2022·Twitter Web App2 Likes