Re: [PATCH v1 7/7] xen-block: implement indirect descriptors

[Jens Axboe]

On Thu, Apr 18 2013, Roger Pau Monné wrote:
> On 18/04/13 14:43, Jens Axboe wrote:
> > On Wed, Apr 17 2013, Konrad Rzeszutek Wilk wrote:
> >> On Wed, Apr 17, 2013 at 07:04:51PM +0200, Roger Pau Monné wrote:
> >>> On 17/04/13 16:25, Konrad Rzeszutek Wilk wrote:
> >>>>>> Perhaps the xen-blkfront part of the patch should be just split out to make
> >>>>>> this easier?
> >>>>>>
> >>>>>> Perhaps what we really should have is just the 'max' value of megabytes
> >>>>>> we want to handle on the ring.
> >>>>>>
> >>>>>> As right now 32 ring requests * 32 segments = 4MB.  But if the user wants
> >>>>>> to se the max: 32 * 4096 = so 512MB (right? each request would handle now 16MB
> >>>>>> and since we have 32 of them = 512MB).
> >>>>>
> >>>>> I've just set that to something that brings a performance benefit
> >>>>> without having to map an insane number of persistent grants in blkback.
> >>>>>
> >>>>> Yes, the values are correct, but the device request queue (rq) is only
> >>>>> able to provide read requests with 64 segments or write requests with
> >>>>> 128 segments. I haven't been able to get larger requests, even when
> >>>>> setting this to 512 or higer.
> >>>>
> >>>> What are you using to drive the requests? 'fio'?
> >>>
> >>> Yes, I've tried fio with several "bs=" values, but it doesn't seem to
> >>> change the size of the underlying requests. Have you been able to get
> >>> bigger requests?
> >>
> >> Martin, Jens,
> >> Any way to drive more than 128 segments?
> > 
> > If the driver is bio based, then there's a natural size constraint on
> > the number of vecs in the bio. So to get truly large requests, the
> > driver would need to merge incoming sequential IOs (similar to how it's
> > done for rq based drivers).
> 
> When you say rq based drivers, you mean drivers with a request queue?
> 
> We are already using a request queue in blkfront, and I'm setting the
> maximum number of segments per request using:
> 
> blk_queue_max_segments(<rq>, <segments>);
> 
> But even when setting <segments> to 256 or 512, I only get read requests
> with 64 segments and write requests with 128 segments from the queue.

What kernel are you testing? The plugging is usually what will trigger a
run of the queue, for rq based drivers. What does your fio job look
like?

-- 
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/