From: Yuanhan Liu <yuanhan.liu@linux.intel.com>
To: Jianbo Liu <jianbo.liu@linaro.org>
Cc: "Wang, Zhihong" <zhihong.wang@intel.com>,
Maxime Coquelin <maxime.coquelin@redhat.com>,
"dev@dpdk.org" <dev@dpdk.org>
Subject: Re: [dpdk-dev] [PATCH v3 0/5] vhost: optimize enqueue
Date: Tue, 27 Sep 2016 18:21:23 +0800 [thread overview]
Message-ID: <20160927102123.GL25823@yliu-dev.sh.intel.com> (raw)
In-Reply-To: <CAP4Qi39YF6SoaiSaka0ioZFWb-2uzWZUbNP4CK7LqCQosaSmWQ@mail.gmail.com>
[-- Attachment #1: Type: text/plain, Size: 2973 bytes --]
On Thu, Sep 22, 2016 at 05:01:41PM +0800, Jianbo Liu wrote:
> On 22 September 2016 at 14:58, Wang, Zhihong <zhihong.wang@intel.com> wrote:
> >
> >
> >> -----Original Message-----
> >> From: Jianbo Liu [mailto:jianbo.liu@linaro.org]
> >> Sent: Thursday, September 22, 2016 1:48 PM
> >> To: Yuanhan Liu <yuanhan.liu@linux.intel.com>
> >> Cc: Wang, Zhihong <zhihong.wang@intel.com>; Maxime Coquelin
> >> <maxime.coquelin@redhat.com>; dev@dpdk.org
> >> Subject: Re: [dpdk-dev] [PATCH v3 0/5] vhost: optimize enqueue
> >>
> >> On 22 September 2016 at 10:29, Yuanhan Liu <yuanhan.liu@linux.intel.com>
> >> wrote:
> >> > On Wed, Sep 21, 2016 at 08:54:11PM +0800, Jianbo Liu wrote:
> >> >> >> > My setup consists of one host running a guest.
> >> >> >> > The guest generates as much 64bytes packets as possible using
> >> >> >>
> >> >> >> Have you tested with other different packet size?
> >> >> >> My testing shows that performance is dropping when packet size is
> >> more
> >> >> >> than 256.
> >> >> >
> >> >> >
> >> >> > Hi Jianbo,
> >> >> >
> >> >> > Thanks for reporting this.
> >> >> >
> >> >> > 1. Are you running the vector frontend with mrg_rxbuf=off?
> >> >> >
> >> Yes, my testing is mrg_rxbuf=off, but not vector frontend PMD.
> >>
> >> >> > 2. Could you please specify what CPU you're running? Is it Haswell
> >> >> > or Ivy Bridge?
> >> >> >
> >> It's an ARM server.
> >>
> >> >> > 3. How many percentage of drop are you seeing?
> >> The testing result:
> >> size (bytes) improvement (%)
> >> 64 3.92
> >> 128 11.51
> >> 256 24.16
> >> 512 -13.79
> >> 1024 -22.51
> >> 1500 -12.22
> >> A correction is that performance is dropping if byte size is larger than 512.
> >
> >
> > Jianbo,
> >
> > Could you please verify does this patch really cause enqueue perf to drop?
> >
> > You can test the enqueue path only by set guest to do rxonly, and compare
> > the mpps by show port stats all in the guest.
> >
> >
> Tested with testpmd, host: txonly, guest: rxonly
> size (bytes) improvement (%)
> 64 4.12
> 128 6
> 256 2.65
> 512 -1.12
> 1024 -7.02
There is a difference between Zhihong's code and the old I spotted in
the first time: Zhihong removed the avail_idx prefetch. I understand
the prefetch becomes a bit tricky when mrg-rx code path is considered;
thus, I didn't comment on that.
That's one of the difference that, IMO, could drop a regression. I then
finally got a chance to add it back.
A rough test shows it improves the performance of 1400B packet size greatly
in the "txonly in host and rxonly in guest" case: +33% is the number I get
with my test server (Ivybridge).
I guess this might/would help your case as well. Mind to have a test
and tell me the results?
BTW, I made it in rush; I haven't tested the mrg-rx code path yet.
Thanks.
--yliu
[-- Attachment #2: diff --]
[-- Type: text/plain, Size: 4341 bytes --]
commit e5852d04bf87c02d6d0d8e6d8ded4c33030b9c9e
Author: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Date: Tue Sep 27 17:51:15 2016 +0800
xxxx
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 381dc27..41bfeba 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -61,6 +61,8 @@ struct buf_vector {
uint32_t desc_idx;
};
+#define NR_AVAIL_IDX_PREFETCH 32
+
/**
* Structure contains variables relevant to RX/TX virtqueues.
*/
@@ -70,7 +72,7 @@ struct vhost_virtqueue {
struct vring_used *used;
uint32_t size;
- /* Last index used on the available ring */
+ uint16_t last_avail_idx;
uint16_t last_used_idx;
#define VIRTIO_INVALID_EVENTFD (-1)
#define VIRTIO_UNINITIALIZED_EVENTFD (-2)
@@ -89,6 +91,9 @@ struct vhost_virtqueue {
/* Shadow used ring for performance */
struct vring_used_elem *shadow_used_ring;
uint32_t shadow_used_idx;
+
+ uint16_t next_avail_idx;
+ uint16_t avail_idx_buf[NR_AVAIL_IDX_PREFETCH];
} __rte_cache_aligned;
/* Old kernels have no such macro defined */
diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 11a2c1a..1cc22fc 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -170,6 +170,41 @@ flush_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
}
}
+/* Fetch NR_AVAIL_IDX_PREFETCH avail entries at once */
+static void
+prefetch_avail_idx(struct vhost_virtqueue *vq)
+{
+ int i;
+
+ for (i = 0; i < NR_AVAIL_IDX_PREFETCH; i++) {
+ vq->avail_idx_buf[i] = vq->avail->ring[(vq->last_avail_idx + i) &
+ (vq->size - 1)];
+ }
+}
+
+static uint16_t
+next_avail_idx(struct vhost_virtqueue *vq)
+{
+ if (vq->next_avail_idx >= NR_AVAIL_IDX_PREFETCH) {
+ prefetch_avail_idx(vq);
+ vq->next_avail_idx = 0;
+ vq->last_avail_idx += NR_AVAIL_IDX_PREFETCH;
+ }
+
+ return vq->avail_idx_buf[vq->next_avail_idx++];
+}
+
+/*
+ * Just peek, but don't move forward the "next_avail_idx" pointer
+ * The caller also has to make sure the point doesn't go beyond
+ * the array.
+ */
+static uint16_t
+peek_next_avail_idx(struct vhost_virtqueue *vq)
+{
+ return vq->avail_idx_buf[vq->next_avail_idx];
+}
+
static inline int __attribute__((always_inline))
enqueue_packet(struct virtio_net *dev, struct vhost_virtqueue *vq,
uint16_t avail_idx, struct rte_mbuf *mbuf,
@@ -193,7 +228,7 @@ enqueue_packet(struct virtio_net *dev, struct vhost_virtqueue *vq,
mbuf_avail = mbuf_len;
/* get the current desc */
- desc_current = vq->avail->ring[(vq->last_used_idx) & (vq->size - 1)];
+ desc_current = next_avail_idx(vq);
desc_chain_head = desc_current;
desc = &vq->desc[desc_current];
desc_addr = gpa_to_vva(dev, desc->addr);
@@ -235,9 +270,7 @@ enqueue_packet(struct virtio_net *dev, struct vhost_virtqueue *vq,
if (avail_idx == vq->last_used_idx)
goto error;
- desc_current =
- vq->avail->ring[(vq->last_used_idx) &
- (vq->size - 1)];
+ desc_current = next_avail_idx(vq);
desc_chain_head = desc_current;
desc_chain_len = 0;
} else
@@ -298,6 +331,7 @@ notify_guest(struct virtio_net *dev, struct vhost_virtqueue *vq)
eventfd_write(vq->callfd, (eventfd_t)1);
}
+
uint16_t
rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
struct rte_mbuf **pkts, uint16_t count)
@@ -331,14 +365,15 @@ rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
/* start enqueuing packets 1 by 1 */
vq->shadow_used_idx = 0;
+ vq->next_avail_idx = 0;
used_idx = vq->last_used_idx & (vq->size - 1);
avail_idx = *((volatile uint16_t *)&vq->avail->idx);
while (pkt_left && avail_idx != vq->last_used_idx) {
/* prefetch the next desc */
- if (pkt_left > 1 && avail_idx != vq->last_used_idx + 1)
- rte_prefetch0(&vq->desc[vq->avail->ring[
- (vq->last_used_idx + 1) &
- (vq->size - 1)]]);
+ if (pkt_left > 1 &&
+ vq->next_avail_idx + 1 < NR_AVAIL_IDX_PREFETCH) {
+ rte_prefetch0(&vq->desc[peek_next_avail_idx(vq)]);
+ }
if (enqueue_packet(dev, vq, avail_idx, pkts[pkt_idx],
is_mrg_rxbuf))
@@ -347,6 +382,7 @@ rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
pkt_idx++;
pkt_left--;
}
+ vq->last_avail_idx += vq->next_avail_idx;
/* batch update used ring for better performance */
if (likely(vq->shadow_used_idx > 0))
next prev parent reply other threads:[~2016-09-27 10:20 UTC|newest]
Thread overview: 140+ messages / expand[flat|nested] mbox.gz Atom feed top
2016-08-16 3:50 [dpdk-dev] [PATCH] optimize vhost enqueue Zhihong Wang
2016-08-16 13:59 ` Maxime Coquelin
2016-08-17 1:45 ` Wang, Zhihong
2016-08-17 2:38 ` Yuanhan Liu
2016-08-17 6:41 ` Wang, Zhihong
2016-08-17 9:17 ` Maxime Coquelin
2016-08-17 9:51 ` Yuanhan Liu
2016-08-18 13:44 ` Wang, Zhihong
2016-08-17 10:07 ` Wang, Zhihong
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 0/6] vhost: optimize enqueue Zhihong Wang
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 1/6] vhost: rewrite enqueue Zhihong Wang
2016-08-19 2:39 ` Yuanhan Liu
2016-08-19 7:07 ` Wang, Zhihong
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 2/6] vhost: remove obsolete Zhihong Wang
2016-08-19 2:32 ` Yuanhan Liu
2016-08-19 7:08 ` Wang, Zhihong
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 3/6] vhost: remove useless volatile Zhihong Wang
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 4/6] vhost: add desc prefetch Zhihong Wang
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 5/6] vhost: batch update used ring Zhihong Wang
2016-08-18 6:33 ` [dpdk-dev] [PATCH v2 6/6] vhost: optimize cache access Zhihong Wang
2016-08-19 5:43 ` [dpdk-dev] [PATCH v3 0/5] vhost: optimize enqueue Zhihong Wang
2016-08-19 5:43 ` [dpdk-dev] [PATCH v3 1/5] vhost: rewrite enqueue Zhihong Wang
2016-08-22 9:35 ` Maxime Coquelin
2016-08-23 2:27 ` Wang, Zhihong
2016-08-25 4:00 ` Yuanhan Liu
2016-08-19 5:43 ` [dpdk-dev] [PATCH v3 2/5] vhost: remove useless volatile Zhihong Wang
2016-08-19 5:43 ` [dpdk-dev] [PATCH v3 3/5] vhost: add desc prefetch Zhihong Wang
2016-08-19 5:43 ` [dpdk-dev] [PATCH v3 4/5] vhost: batch update used ring Zhihong Wang
2016-08-25 3:48 ` Yuanhan Liu
2016-08-25 5:19 ` Wang, Zhihong
2016-08-19 5:43 ` [dpdk-dev] [PATCH v3 5/5] vhost: optimize cache access Zhihong Wang
2016-08-22 8:11 ` [dpdk-dev] [PATCH v3 0/5] vhost: optimize enqueue Maxime Coquelin
2016-08-22 10:01 ` Maxime Coquelin
2016-08-22 10:35 ` Thomas Monjalon
2016-08-24 3:37 ` Wang, Zhihong
2016-08-23 2:31 ` Wang, Zhihong
2016-08-23 10:43 ` Wang, Zhihong
2016-08-23 12:16 ` Maxime Coquelin
2016-08-23 12:22 ` Yuanhan Liu
2016-08-23 2:15 ` Wang, Zhihong
2016-09-21 8:50 ` Jianbo Liu
2016-09-21 9:27 ` Wang, Zhihong
2016-09-21 12:54 ` Jianbo Liu
2016-09-22 2:11 ` Wang, Zhihong
2016-09-22 2:29 ` Yuanhan Liu
2016-09-22 5:47 ` Jianbo Liu
2016-09-22 6:58 ` Wang, Zhihong
2016-09-22 9:01 ` Jianbo Liu
2016-09-22 10:04 ` Wang, Zhihong
2016-09-22 14:41 ` Jianbo Liu
2016-09-23 2:56 ` Wang, Zhihong
2016-09-23 10:41 ` Jianbo Liu
2016-09-23 13:41 ` Thomas Monjalon
2016-09-25 5:41 ` Wang, Zhihong
2016-09-26 5:12 ` Jianbo Liu
2016-09-26 5:25 ` Wang, Zhihong
2016-09-26 5:38 ` Jianbo Liu
2016-09-26 6:00 ` Wang, Zhihong
2016-09-26 4:24 ` Jianbo Liu
2016-09-26 5:37 ` Luke Gorrie
2016-09-26 5:40 ` Jianbo Liu
2016-09-27 10:21 ` Yuanhan Liu [this message]
2016-09-27 16:45 ` Wang, Zhihong
2016-10-09 12:09 ` Wang, Zhihong
2016-10-10 2:44 ` Yuanhan Liu
2016-10-10 5:31 ` Jianbo Liu
2016-10-10 6:22 ` Wang, Zhihong
2016-10-10 6:57 ` Jianbo Liu
2016-10-10 7:25 ` Wang, Zhihong
2016-10-12 2:53 ` Yuanhan Liu
2016-10-12 12:22 ` Wang, Zhihong
2016-10-12 15:31 ` Thomas Monjalon
2016-10-13 1:21 ` Wang, Zhihong
2016-10-13 3:51 ` Jianbo Liu
2016-10-13 5:33 ` Yuanhan Liu
2016-10-13 5:35 ` Yuanhan Liu
2016-10-13 6:02 ` Wang, Zhihong
2016-10-13 7:54 ` Maxime Coquelin
2016-10-13 9:23 ` Maxime Coquelin
2016-10-14 10:11 ` Yuanhan Liu
2016-08-30 3:35 ` [dpdk-dev] [PATCH v4 0/6] " Zhihong Wang
2016-08-30 3:35 ` [dpdk-dev] [PATCH v4 1/6] vhost: fix windows vm hang Zhihong Wang
2016-09-05 5:24 ` [dpdk-dev] [dpdk-stable] " Yuanhan Liu
2016-09-05 5:25 ` Wang, Zhihong
2016-09-05 5:40 ` Yuanhan Liu
2016-08-30 3:36 ` [dpdk-dev] [PATCH v4 2/6] vhost: rewrite enqueue Zhihong Wang
2016-09-05 6:39 ` Yuanhan Liu
2016-09-07 5:33 ` Yuanhan Liu
2016-09-07 5:39 ` Wang, Zhihong
2016-08-30 3:36 ` [dpdk-dev] [PATCH v4 3/6] vhost: remove useless volatile Zhihong Wang
2016-08-30 3:36 ` [dpdk-dev] [PATCH v4 4/6] vhost: add desc prefetch Zhihong Wang
2016-08-30 3:36 ` [dpdk-dev] [PATCH v4 5/6] vhost: batch update used ring Zhihong Wang
2016-08-30 3:36 ` [dpdk-dev] [PATCH v4 6/6] vhost: optimize cache access Zhihong Wang
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 0/6] vhost: optimize enqueue Zhihong Wang
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 1/6] vhost: fix windows vm hang Zhihong Wang
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 2/6] vhost: rewrite enqueue Zhihong Wang
2016-09-12 15:42 ` Maxime Coquelin
2016-09-14 8:20 ` Wang, Zhihong
2016-09-15 16:35 ` Maxime Coquelin
2016-09-12 16:26 ` Maxime Coquelin
2016-09-14 8:22 ` Wang, Zhihong
2016-09-18 14:19 ` Yuanhan Liu
2016-09-19 3:29 ` Wang, Zhihong
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 3/6] vhost: remove useless volatile Zhihong Wang
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 4/6] vhost: add desc prefetch Zhihong Wang
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 5/6] vhost: batch update used ring Zhihong Wang
2016-09-12 15:45 ` Maxime Coquelin
2016-09-14 8:43 ` Wang, Zhihong
2016-09-15 16:38 ` Maxime Coquelin
2016-09-18 2:55 ` Yuanhan Liu
2016-09-18 2:57 ` Wang, Zhihong
2016-09-09 3:39 ` [dpdk-dev] [PATCH v5 6/6] vhost: optimize cache access Zhihong Wang
2016-09-12 13:52 ` [dpdk-dev] [PATCH v5 0/6] vhost: optimize enqueue Maxime Coquelin
2016-09-12 13:56 ` Maxime Coquelin
2016-09-12 14:01 ` Yuanhan Liu
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 " Zhihong Wang
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 1/6] vhost: fix windows vm hang Zhihong Wang
2016-10-13 6:18 ` [dpdk-dev] [dpdk-stable] " Yuanhan Liu
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 2/6] vhost: rewrite enqueue Zhihong Wang
2016-09-22 9:58 ` Jianbo Liu
2016-09-22 10:13 ` Wang, Zhihong
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 3/6] vhost: remove useless volatile Zhihong Wang
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 4/6] vhost: add desc prefetch Zhihong Wang
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 5/6] vhost: batch update used ring Zhihong Wang
2016-09-20 2:00 ` [dpdk-dev] [PATCH v6 6/6] vhost: optimize cache access Zhihong Wang
2016-09-21 2:26 ` [dpdk-dev] [PATCH v6 0/6] vhost: optimize enqueue Yuanhan Liu
2016-09-21 4:39 ` Maxime Coquelin
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 0/7] vhost: optimize mergeable Rx path Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 1/7] vhost: remove useless volatile Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 2/7] vhost: optimize cache access Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 3/7] vhost: simplify mergeable Rx vring reservation Yuanhan Liu
2016-10-25 22:08 ` Thomas Monjalon
2016-10-26 2:56 ` Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 4/7] vhost: use last avail idx for avail ring reservation Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 5/7] vhost: shadow used ring update Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 6/7] vhost: prefetch avail ring Yuanhan Liu
2016-10-14 9:34 ` [dpdk-dev] [PATCH v7 7/7] vhost: retrieve avail head once Yuanhan Liu
2016-10-18 2:25 ` [dpdk-dev] [PATCH v7 0/7] vhost: optimize mergeable Rx path Jianbo Liu
2016-10-18 14:53 ` Maxime Coquelin
2016-10-21 7:51 ` Yuanhan Liu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20160927102123.GL25823@yliu-dev.sh.intel.com \
--to=yuanhan.liu@linux.intel.com \
--cc=dev@dpdk.org \
--cc=jianbo.liu@linaro.org \
--cc=maxime.coquelin@redhat.com \
--cc=zhihong.wang@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).