On Thu, 2012-09-20 at 22:24 +0100, Konrad Rzeszutek Wilk wrote:
> On Thu, Sep 20, 2012 at 03:13:42PM +0100, Oliver Chick wrote:
> > On Thu, 2012-09-20 at 14:49 +0100, Konrad Rzeszutek Wilk wrote:
> > > On Thu, Sep 20, 2012 at 12:48:41PM +0100, Jan Beulich wrote:
> > > > >>> On 20.09.12 at 13:30, Oliver Chick <oliver.ch...@citrix.com> wrote:
> > > > > The memory overhead, and fallback mode points are related:
> > > > > -Firstly, it turns out that the overhead is actually 2.75MB, not 11MB
> > > > > per device. I made a mistake (pointed out by Jan) as the maximum
> > > > > number
> > > > > of requests that can fit into a single-page ring is 64, not 256.
> > > > > -Clearly, this still scales linearly. So the problem of memory
> > > > > footprint
> > > > > will occur with more VMs, or block devices.
> > > > > -Whilst 2.75MB per device is probably acceptable (?), if we start
> > > > > using
> > > > > multipage rings, then we might not want to have
> > > > > BLKIF_MAX_PERS_REQUESTS_PER_DEVICE==__RING_SIZE, as this will cause
> > > > > the
> > > > > memory overhead to increase. This is why I have implemented the
> > > > > 'fallback' mode. With a multipage ring, it seems reasonable to want
> > > > > the
> > > > > first $x$ grefs seen by blkback to be treated as persistent, and any
> > > > > later ones to be non-persistent. Does that seem sensible?
> > > >
> > > > From a resource usage pov, perhaps. But this will get the guest
> > > > entirely unpredictable performance. Plus I don't think 11Mb of
> > >
> > > Wouldn't it fall back to the older performance?
> >
> > I guess it would be a bit more complex than that. It would be worse than
> > the new performance because the grefs that get processed by the
> > 'fallback' mode will cause TLB shootdowns. But any early grefs will
> > still be processed by the persistent mode, so won't have shootdowns.
> > Therefore, depending on the ratio of {persistent grants}:{non-persistent
> > grants), allocated by blkfront, the performance will be somewhere
> > inbetween the two extremes.
> >
> > I guess that the choice is between
> > 1) Compiling blk{front,back} with a pre-determined number of persistent
> > grants, and failing if this limit is exceeded. This seems rather
> > unflexible, as blk{front,back} must then both both use the same version,
> > or you will get failures.
> > 2 (current setup)) Have a recommended maximum number of
> > persistently-mapped pages, and going into a 'fallback' mode if blkfront
> > exceeds this limit.
> > 3) Having blkback inform blkfront on startup as to how many grefs it is
> > willing to persistently-map. We then hit the same question again though:
> > what should be do if blkfront ignores this limit?
>
> How about 2 and 3 together?
I think 1 is fine for a "phase 1" implementation, especially taking into
consideration that the end of Oliver's internship is next week.
Also it seems that the cases where there might be some disconnect
between the number of persistent grants supported by the backend and the
number of requests from the frontend are currently theoretical or
predicated on the existence of unmerged or as yet unwritten patches.
So lets say, for now, that the default number of persistent grants is
the same as the number of slots in the ring and that it is a bug for
netfront to try and use more than that if it has signed up to the use of
persistent grants. netback is at liberty to fail such "overflow"
requests. In practice this can't happen with the current implementations
given the default specified above.
If someone wants to implement something like 2 or 3 in the future then
they can do so by negotiating through a xenstore key for a non-default
number of pgrants.
I think that if/when the number of persistent grants can differ from the
number of ring slots that an LRU scheme would be best. i.e. if there are
N slots then when the N+1th unique grant comes in we discard the least
recently used N/M (M perhaps in {2,3,4}) of the persistently granted
pages. This way there is an incentive for the f.e. to try to reuse pages
as much as possible and we get good batching on the unmaps if not.
Ian.
> Meaning have a recommended maximmum number.
> If we fall back due to memory pressure we can tell the guest that we
> are entering fall-back mode. The frontend can decide what it wants to do
> (throttle the amount of I/Os?) or just do a printk telling the user it
> dropped the speed from "Insane Hot!" down to "Turbo!"...
>
> Or maybe not. Perhaps just reporting it in the backend that we are
> hitting memory pressure and using the old-style-fallback mechanism
> so the system admin can take actions (and tell his users why suddenly
> their I/Os are so slow).
>
> _______________________________________________
> Xen-devel mailing list
> xen-de...@lists.xen.org
> http://lists.xen.org/xen-devel
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
