[dm-devel] scheduling/merging on top of a DM target?

[Ming Lei]

On Mon, Oct 23, 2017 at 2:16 AM, Peter Desnoyers <pjd at ccs.neu.edu> wrote:
> We’re emulating some in-disk algorithms in a device mapper target, and would like to be able to do a fair comparison between the emulated and real versions; however they see radically different I/O streams, evidently because there’s no scheduling / request merging / etc. on top of the bio-based device mapper target.
>
> It appears that if our device mapper were request-based instead of bio-based we wouldn’t have this issue, but after reading a bunch of history on the mailing list archives I don’t think this is going to happen easily.
>
> Does anyone have any thoughts on quick-and-dirty ways to approximate this so we can get some preliminary measurements? dm-multipath won’t stack on a bio-based target, and attempts to stack loopback and various iscsi targets on top haven’t achieved the desired results. (in general these have resulted in our dm target seeing individual 4K I/Os, instead of the larger I/Os at the higher layer)
>

I posted out one patchset for addressing the issue:

https://marc.info/?l=linux-block&m=150677334021817&w=2

But it has been obsolete, will post out a new version
when I get time to do that.


-- 
Ming Lei