I would like to increase the performance of the multibus mode of multipath.
It would be nice if multibus performance would be very close to accessing
the device directly. From what I have been told, under multibus, I/O's are
broken up into bio's and sent down the multiple paths, with no merging of
adjacent I/O's. Would it be possible to add code to the kernel so that
adjacent (at least some of them, probably make configurable) would be merged
and sent on a single path? And, would this increase the performance of
multibus mode?

I would be willing to work on it myself, but I have never done any kernel
coding (I done plenty of C coding) and would not know where to look to even
attempt adding I/O merging to multipath.

What do you think? Would this work, and how would I go about doing it?

Andy

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel

Andy [genanr@emsphone.com] wrote:
> I would like to increase the performance of the multibus mode of multipath.
> It would be nice if multibus performance would be very close to accessing
> the device directly. From what I have been told, under multibus, I/O's are
> broken up into bio's and sent down the multiple paths, with no merging of
> adjacent I/O's.

Not quite correct. It depends on the path selector algorithm you use.
"round_robin" with rr_min_io set to 1 will definitely give you bad
performance for sequential I/O. If you set rr_min_io to a big number,
the adjacent I/Os may be merged together. What values did you use in
your testing?

--Malahal.

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel