Re: [dm-devel] [PATCH] dm mpath selector: more evenly distribute ties

[Khazhismel Kumykov]

On Wed, Jan 24, 2018 at 11:09 AM, Martin Wilck <mwilck@xxxxxxxx> wrote:
> On Wed, 2018-01-24 at 10:44 -0800, Khazhismel Kumykov wrote:
>> On Wed, Jan 24, 2018 at 2:57 AM, Martin Wilck <mwilck@xxxxxxxx>
>> wrote:
>> > On Fri, 2018-01-19 at 15:07 -0800, Khazhismel Kumykov wrote:
>> > > Move the last used path to the end of the list (least preferred)
>> > > so
>> > > that
>> > > ties are more evenly distributed.
>> > >
>> > > For example, in case with three paths with one that is slower
>> > > than
>> > > others, the remaining two would be unevenly used if they tie.
>> > > This is
>> > > due to the rotation not being a truely fair distribution.
>> > >
>> > > Illustrated: paths a, b, c, 'c' has 1 outstanding IO, a and b are
>> > > 'tied'
>> > > Three possible rotations:
>> > > (a, b, c) -> best path 'a'
>> > > (b, c, a) -> best path 'b'
>> > > (c, a, b) -> best path 'a'
>> > > (a, b, c) -> best path 'a'
>> > > (b, c, a) -> best path 'b'
>> > > (c, a, b) -> best path 'a'
>> > > ...
>> >
>> >
>> > This happens only if a and b actually have the same weight (e.g.
>> > queue
>> > length for the queue-length selector). If 'a' really receives more
>> > IO,
>> > its queue grows, and the selector will start preferring 'b', so the
>> > effect should level out automatically with the current code as soon
>> > as
>> > you have real IO going on. But maybe I haven't grasped what you're
>> > referring to as "tied".
>>
>> Yes, for "tied" I'm referring to paths with the same weight. As queue
>> length grows it does tend to level out, but in the case where queue
>> length doesn't grow (in this example I'm imagining 2 concurrent
>> requests to the device) the bias does show and the selectors really
>> do
>> send 'a' 2x more requests than 'b' (when 'c' is much slower and 'a'
>> and 'b' are ~same speed).
>
> Have you actually observed this? I find the idea pretty academical that
> two paths would be walking "tied" this way. In practice, under IO load,
> I'd expect queue lengths (and service-times, for that matter) to
> fluctuate enough to prevent this effect to be measurable. But of
> course, I may be wrong. If you really did observe this, the next
> question would be: does hurt performance to an extent that can be
> noticed/measured? I reckon that if 'a' got saturated under the load,
> hurting performance, its queue length would grow quickly and 'b' would
> automatically get preferred.
This is fairly simple to observe - start two sync reader threads
against a device with 3 backing paths, with one path taking longer on
average to complete requests than the other two. One of the 'faster'
paths will be used ~2x more than the other. Perhaps not that common a
situation, but is a real one. The bias seemed simple to remove, so
that the two (or N) paths would be used equally.

I don't see a downside to more evenly distributing in this case,
although I can't say I've directly observed performance downsides for
a biased distribution (aside from the distribution being biased itself
being a downside).
The bias of note is inherent to the order paths are added to the
selector (and which path is 'always bad'), so if 'a' is saturated due
to this, then slows, once it recovers it will continue to be
preferred, versus in an even distribution.

Khazhy

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature