On Fri, 26 Jan 2024, Hannes Reinecke wrote: > Oh, it's this time of the year again? > (This topic regularly comes up ...) > > The reason is not that it will disable dm-multipath (Mike Snitzer put in > bio-based multipathing as an additional code path); the reason is that > dm-multipath performance will suffer when you remove request-based DM. Is there some benchmark that says how much will it suffer? > DM-multipath schedules based on request (if you use the request-based > interface) or bios (if you use the bio-based interface). > Any merge decision is typically done by the block layer when combining bios > into requests; and you can only merge bios if the bvecs are adjacent. > So if you use bio-based multipathing you will spread sequential bios > across all paths, leaving the block layer unable to merge requests. The same problem exists in raid1 and there's a function read_balance that solves it. If the starting sector of a new bio matches the ending sector of a previous bio, then submit it for the same device. > For request based multipathing the requests are already fully-formed, > and scheduling across paths does not change them. > Things are slightly better with multi-page bvecs nowadays, but the > overall picture still stands. > > Another thing is timeouts; bios don't do timeouts, so a bio can run > for an arbitrary time with no chance of interrupting it. > Requests do have a timeout, and will be aborted from the driver when > the timeout is hit. > Seeing that 99% of all I/O issues I've seen _are_ timeouts it becomes > a crucial feature if you want dm-multipath to control failover time. You can set timeout of the underlying physical devices. > Cheers, > > Hannes Mikulas
