Re: [Qemu-devel] Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On 05/31/10 09:46, Rusty Russell wrote: > On Thu, 27 May 2010 05:20:35 am Michael S. Tsirkin wrote: >> Here's a rewrite of the original patch with a new layout. >> I haven't tested it yet so no idea how this performs, but >> I think this addresses the cache bounce issue raised by Avi. >> Posting for early flames/comments. > > Sorry, not without some evidence that it'll actually reduce cacheline > bouncing. I *think* it will, but it's not obvious: the host may keep > looking at avail_idx as we're updating last_seen. Or does qemu always > look at both together anyway? > > Can someone convince me this is a win? > Rusty. Hi Rusty, I ran some tests using the vring index publish patch with virtio_blk. The numbers are based on running IOZone on a ramdisk passed to the guest via virtio. While I didn't see any throughput improvements, I saw a 20-30% reduction in the VMExit count for the full run. This was measured grabbing the VMExit count prior and after the run and calculating the difference. I have the numbers in a PDF, so I will email it to you in private as I don't like sending PDFs to the mailing list. However if anybody else wants the numbers feel free to ping me off list and I'll forward them. Cheers, Jes -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On Mon, May 31, 2010 at 09:36:00AM +0200, Jes Sorensen wrote: > On 05/30/10 13:22, Michael S. Tsirkin wrote: > > On Fri, May 28, 2010 at 11:56:54AM +0200, Jes Sorensen wrote: > >> It looks pretty good to me, however one thing I have been thinking of > >> while reading through it: > >> > >> Rather than storing a pointer within the ring struct, pointing into a > >> position within the same struct. How about storing a byte offset instead > >> and using a cast to get to the pointer position? That would avoid the > >> pointer dereference, which is less effective cache wise and harder for > >> the CPU to predict. > >> > >> Not sure whether it really matters performance wise, just a thought. > > > > I think this won't work: when PUBLUSH_USED_IDX is negotiated, > > the pointer is to within the ring. > > Hmmm shame, it would be a nice optimization. > > Maybe it's time to introduce the v2 ring format, rather than having > adding more kludges to the existing one? > > Cheers, > Jes There has been discussion about a ring format that does not use indexes at all. My guess is that would be a good point for v2 ring format. But making that a product and tuning might take a while. So definitely something to keep in mind but I would not want that to block this optimization. -- MST -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On Mon, May 31, 2010 at 05:16:42PM +0930, Rusty Russell wrote: > On Thu, 27 May 2010 05:20:35 am Michael S. Tsirkin wrote: > > Here's a rewrite of the original patch with a new layout. > > I haven't tested it yet so no idea how this performs, but > > I think this addresses the cache bounce issue raised by Avi. > > Posting for early flames/comments. > > Sorry, not without some evidence that it'll actually reduce cacheline > bouncing. I *think* it will, but it's not obvious: the host may keep > looking at avail_idx as we're updating last_seen. Or does qemu always > look at both together anyway? > Can someone convince me this is a win? > Rusty. What really happens is host looks at flags and last_seen together. And flags happens to be in the same cache line with avail idx. So to get an obvious win, we should put flags and last_seen in a separate cache line from avail, which us easy - just add some padding. And I'll relax the requirement from guest to only require it to update last_seen when interrupts are enabled. This way flags and last_seen are written together and read together. Makes sense? -- MST -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On Thu, 27 May 2010 05:20:35 am Michael S. Tsirkin wrote: > Here's a rewrite of the original patch with a new layout. > I haven't tested it yet so no idea how this performs, but > I think this addresses the cache bounce issue raised by Avi. > Posting for early flames/comments. Sorry, not without some evidence that it'll actually reduce cacheline bouncing. I *think* it will, but it's not obvious: the host may keep looking at avail_idx as we're updating last_seen. Or does qemu always look at both together anyway? Can someone convince me this is a win? Rusty. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On 05/30/10 13:22, Michael S. Tsirkin wrote: > On Fri, May 28, 2010 at 11:56:54AM +0200, Jes Sorensen wrote: >> It looks pretty good to me, however one thing I have been thinking of >> while reading through it: >> >> Rather than storing a pointer within the ring struct, pointing into a >> position within the same struct. How about storing a byte offset instead >> and using a cast to get to the pointer position? That would avoid the >> pointer dereference, which is less effective cache wise and harder for >> the CPU to predict. >> >> Not sure whether it really matters performance wise, just a thought. > > I think this won't work: when PUBLUSH_USED_IDX is negotiated, > the pointer is to within the ring. Hmmm shame, it would be a nice optimization. Maybe it's time to introduce the v2 ring format, rather than having adding more kludges to the existing one? Cheers, Jes -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On Fri, May 28, 2010 at 11:56:54AM +0200, Jes Sorensen wrote: > On 05/26/10 21:50, Michael S. Tsirkin wrote: > > Here's a rewrite of the original patch with a new layout. > > I haven't tested it yet so no idea how this performs, but > > I think this addresses the cache bounce issue raised by Avi. > > Posting for early flames/comments. > > > > Generally, the Host end of the virtio ring doesn't need to see where > > Guest is up to in consuming the ring. However, to completely understand > > what's going on from the outside, this information must be exposed. > > For example, host can reduce the number of interrupts by detecting > > that the guest is currently handling previous buffers. > > > > We add a feature bit so the guest can tell the host that it's writing > > out the current value there, if it wants to use that. > > > > This differs from original approach in that the used index > > is put after avail index (they are typically written out together). > > To avoid cache bounces on descriptor access, > > and make future extensions easier, we put the ring itself at start of > > page, and move the control after it. > > Hi Michael, > > It looks pretty good to me, however one thing I have been thinking of > while reading through it: > > Rather than storing a pointer within the ring struct, pointing into a > position within the same struct. How about storing a byte offset instead > and using a cast to get to the pointer position? That would avoid the > pointer dereference, which is less effective cache wise and harder for > the CPU to predict. > > Not sure whether it really matters performance wise, just a thought. > > Cheers, > Jes I think this won't work: when PUBLUSH_USED_IDX is negotiated, the pointer is to within the ring. -- MST -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On 05/26/10 21:50, Michael S. Tsirkin wrote: > Here's a rewrite of the original patch with a new layout. > I haven't tested it yet so no idea how this performs, but > I think this addresses the cache bounce issue raised by Avi. > Posting for early flames/comments. > > Generally, the Host end of the virtio ring doesn't need to see where > Guest is up to in consuming the ring. However, to completely understand > what's going on from the outside, this information must be exposed. > For example, host can reduce the number of interrupts by detecting > that the guest is currently handling previous buffers. > > We add a feature bit so the guest can tell the host that it's writing > out the current value there, if it wants to use that. > > This differs from original approach in that the used index > is put after avail index (they are typically written out together). > To avoid cache bounces on descriptor access, > and make future extensions easier, we put the ring itself at start of > page, and move the control after it. Hi Michael, It looks pretty good to me, however one thing I have been thinking of while reading through it: Rather than storing a pointer within the ring struct, pointing into a position within the same struct. How about storing a byte offset instead and using a cast to get to the pointer position? That would avoid the pointer dereference, which is less effective cache wise and harder for the CPU to predict. Not sure whether it really matters performance wise, just a thought. Cheers, Jes -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On Thu, May 27, 2010 at 03:07:52PM +0300, Avi Kivity wrote:
> I missed the spec patch, can you repost it?
Still work in progress, but here it is.
Note I am still debating with myself whether we should split
avail idx and flags into separate cache lines.
diff --git a/virtio-spec.lyx b/virtio-spec.lyx
index ed35893..150e5a8 100644
--- a/virtio-spec.lyx
+++ b/virtio-spec.lyx
@@ -1803,6 +1803,36 @@ next
\emph default
descriptor entry (modulo the ring size).
This starts at 0, and increases.
+\change_inserted 0 1274966643
+
+\end_layout
+
+\begin_layout Standard
+
+\change_inserted 0 1274968378
+When PUBLISH_USED feature flag has
+\emph on
+not
+\emph default
+ been negotiated, the ring follows the
+\begin_inset Quotes eld
+\end_inset
+
+flags
+\begin_inset Quotes erd
+\end_inset
+
+ and the
+\begin_inset Quotes eld
+\end_inset
+
+idx
+\begin_inset Quotes erd
+\end_inset
+
+ fields:
+\change_unchanged
+
\end_layout
\begin_layout Standard
@@ -1845,7 +1875,134 @@ struct vring_avail {
\end_layout
+\begin_layout Standard
+
+\change_inserted 0 1274968432
+\begin_inset CommandInset label
+LatexCommand label
+name "PUBLISH_USED-feature"
+
+\end_inset
+
+When PUBLISH_USED feature flag has been negotiated, the control structure
+ including the
+\begin_inset Quotes eld
+\end_inset
+
+flags and the
+\begin_inset Quotes eld
+\end_inset
+
+idx
+\begin_inset Quotes erd
+\end_inset
+
+ fields follows the ring.
+ This leaves the room for the
+\begin_inset Quotes eld
+\end_inset
+
+last_seen_used_idx
+\begin_inset Quotes erd
+\end_inset
+
+ field, which indicates the most recent
+\begin_inset Quotes eld
+\end_inset
+
+idx
+\begin_inset Quotes erd
+\end_inset
+
+ value observed by guest in the used ring (see
+\begin_inset CommandInset ref
+LatexCommand ref
+reference "sub:Used-Ring"
+
+\end_inset
+
+ below):
+\end_layout
+
+\begin_layout Standard
+
+\change_inserted 0 1274967396
+\begin_inset listings
+inline false
+status open
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274967404
+
+struct vring_avail {
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274967405
+
+ u16 ring[qsz]; /* qsz is the Queue Size field read from device */
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274967533
+
+#define VRING_AVAIL_F_NO_INTERRUPT 1
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274967533
+
+ u16 flags;
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274967533
+
+ u16 idx;
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274968345
+
+ u16 last_seen_used_idx;
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274967396
+
+};
+\end_layout
+
+\end_inset
+
+
+\end_layout
+
+\begin_layout Standard
+
+\change_inserted 0 1274967715
+If the ring is large enough, the second layout maintains the control and
+ ring structures on separate cache lines.
+\end_layout
+
\begin_layout Subsection
+
+\change_inserted 0 1274968415
+\begin_inset CommandInset label
+LatexCommand label
+name "sub:Used-Ring"
+
+\end_inset
+
+
+\change_unchanged
Used Ring
\end_layout
@@ -2391,12 +2548,20 @@ status open
\begin_layout Plain Layout
-while (vq->last_seen_used != vring->used.idx) {
+while (vq->last_seen_used
+\change_inserted 0 1274968316
+_idx
+\change_unchanged
+ != vring->used.idx) {
\end_layout
\begin_layout Plain Layout
-struct vring_used_elem *e = vring.used->ring[vq->last_seen_used%vsz];
+struct vring_used_elem *e = vring.used->ring[vq->last_seen_used
+\change_inserted 0 1274968326
+_idx
+\change_unchanged
+%vsz];
\end_layout
\begin_layout Plain Layout
@@ -2406,7 +2571,11 @@ while (vq->last_seen_used != vring->used.idx) {
\begin_layout Plain Layout
-vq->last_seen_used++;
+vq->last_seen_used
+\change_inserted 0 1274968321
+_idx
+\change_unchanged
+++;
\end_layout
\begin_layout Plain Layout
@@ -2419,6 +2588,13 @@ while (vq->last_seen_used != vring->used.idx) {
\end_layout
+\begin_layout Standard
+
+\change_inserted 0 1274968252
+If PUBLISH_USED feature is negotiated, last_seen_used value should be published
+ to the device in the avail ring.
+\end_layout
+
\begin_layout Subsection
Dealing With Configuration Changes
\end_layout
@@ -2986,6 +3162,47 @@ struct vring_avail {
\begin_layout Plain Layout
};
+\change_inserted 0 1274966477
+
+\end_layout
+
+\begin_layout Plain Layout
+
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274966484
+
+struct vring_avail_ctrl {
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274966489
+
+__u16 flags;
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274966494
+
+__u16 idx;
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274966499
+
+__u16 last_used_idx;
+\end_layout
+
+\begin_layout Plain Layout
+
+\change_inserted 0 1274966474
+
+};
\end_layout
\begin_layout Plain Layout
@@ -3349,6 +3566,28 @@ reference "sub:Indirect-Descript
Re: [PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
On 05/26/2010 10:50 PM, Michael S. Tsirkin wrote: Here's a rewrite of the original patch with a new layout. I haven't tested it yet so no idea how this performs, but I think this addresses the cache bounce issue raised by Avi. Posting for early flames/comments. Generally, the Host end of the virtio ring doesn't need to see where Guest is up to in consuming the ring. However, to completely understand what's going on from the outside, this information must be exposed. For example, host can reduce the number of interrupts by detecting that the guest is currently handling previous buffers. We add a feature bit so the guest can tell the host that it's writing out the current value there, if it wants to use that. This differs from original approach in that the used index is put after avail index (they are typically written out together). To avoid cache bounces on descriptor access, and make future extensions easier, we put the ring itself at start of page, and move the control after it. I missed the spec patch, can you repost it? -- error compiling committee.c: too many arguments to function -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
[PATCHv2-RFC 0/2] virtio: put last seen used index into ring itself
Here's a rewrite of the original patch with a new layout. I haven't tested it yet so no idea how this performs, but I think this addresses the cache bounce issue raised by Avi. Posting for early flames/comments. Generally, the Host end of the virtio ring doesn't need to see where Guest is up to in consuming the ring. However, to completely understand what's going on from the outside, this information must be exposed. For example, host can reduce the number of interrupts by detecting that the guest is currently handling previous buffers. We add a feature bit so the guest can tell the host that it's writing out the current value there, if it wants to use that. This differs from original approach in that the used index is put after avail index (they are typically written out together). To avoid cache bounces on descriptor access, and make future extensions easier, we put the ring itself at start of page, and move the control after it. Signed-off-by: Michael S. Tsirkin Michael S. Tsirkin (2): virtio: support layout with avail ring before idx virtio: publish used idx drivers/net/virtio_net.c |2 ++ drivers/virtio/virtio_ring.c | 19 --- include/linux/virtio_ring.h | 34 -- 3 files changed, 46 insertions(+), 9 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
