From: Stefano Brivio <sbrivio@redhat.com>
To: David Gibson <david@gibson.dropbear.id.au>
Cc: passt-dev@passt.top
Subject: Re: [PATCH 4/8] udp: Receive multiple datagrams at once on the pasta sock->tap path
Date: Wed, 4 Jan 2023 01:08:52 +0100 [thread overview]
Message-ID: <20230104010852.02e96a70@elisabeth> (raw)
In-Reply-To: <Y6Kg+DJ3YR7K2LXR@yekko>
On Wed, 21 Dec 2022 17:00:24 +1100
David Gibson <david@gibson.dropbear.id.au> wrote:
> On Tue, Dec 20, 2022 at 11:42:46AM +0100, Stefano Brivio wrote:
> > Sorry for the further delay,
> >
> > On Wed, 14 Dec 2022 11:35:46 +0100
> > Stefano Brivio <sbrivio@redhat.com> wrote:
> >
> > > On Wed, 14 Dec 2022 12:42:14 +1100
> > > David Gibson <david@gibson.dropbear.id.au> wrote:
> > >
> > > > On Tue, Dec 13, 2022 at 11:48:47PM +0100, Stefano Brivio wrote:
> > > > > Sorry for the long delay here,
> > > > >
> > > > > On Mon, 5 Dec 2022 19:14:21 +1100
> > > > > David Gibson <david@gibson.dropbear.id.au> wrote:
> > > > >
> > > > > > Usually udp_sock_handler() will receive and forward multiple (up to 32)
> > > > > > datagrams in udp_sock_handler(), then forward them all to the tap
> > > > > > interface. For unclear reasons, though, when in pasta mode we will only
> > > > > > receive and forward a single datagram at a time. Change it to receive
> > > > > > multiple datagrams at once, like the other paths.
> > > > >
> > > > > This is explained in the commit message of 6c931118643c ("tcp, udp:
> > > > > Receive batching doesn't pay off when writing single frames to tap").
> > > > >
> > > > > I think it's worth re-checking the throughput now as this path is a bit
> > > > > different, but unfortunately I didn't include this in the "perf" tests :(
> > > > > because at the time I introduced those I wasn't sure it even made sense to
> > > > > have traffic from the same host being directed to the tap device.
> > > > >
> > > > > The iperf3 runs were I observed this are actually the ones from the Podman
> > > > > demo. Ideally that case should be also checked in the perf/pasta_udp tests.
> > > >
> > > > Hm, ok.
> > > >
> > > > > How fundamental is this for the rest of the series? I couldn't find any
> > > > > actual dependency on this but I might be missing something.
> > > >
> > > > So the issue is that prior to this change in pasta we receive multiple
> > > > frames at once on the splice path, but one frame at a time on the tap
> > > > path. By the end of this series we can't do that any more, because we
> > > > don't know before the recvmmsg() which one we'll be doing.
> > >
> > > Oh, right, I see. Then let me add this path to the perf/pasta_udp test
> > > and check how relevant this is now, I'll get back to you in a bit.
> >
> > I was checking the wrong path. With this:
> >
> > diff --git a/test/perf/pasta_udp b/test/perf/pasta_udp
> > index 27ea724..973c2f4 100644
> > --- a/test/perf/pasta_udp
> > +++ b/test/perf/pasta_udp
> > @@ -31,6 +31,14 @@ report pasta lo_udp 1 __FREQ__
> >
> > th MTU 1500B 4000B 16384B 65535B
> >
> > +tr UDP throughput over IPv6: host to ns
> > +nsout IFNAME ip -j link show | jq -rM '.[] | select(.link_type == "ether").ifname'
> > +nsout ADDR6 ip -j -6 addr show|jq -rM '.[] | select(.ifname == "__IFNAME__").addr_info[] | select(.scope == "global" and .prefixlen == 64).local'
> > +bw -
> > +bw -
> > +bw -
> > +iperf3 BW host ns __ADDR6__ 100${i}2 __THREADS__ __TIME__ __OPTS__ -b 15G
> > +bw __BW__ 7.0 9.0
> >
> > tr UDP throughput over IPv6: ns to host
> > ns ip link set dev lo mtu 1500
> > diff --git a/test/run b/test/run
> > index e07513f..b53182b 100755
> > --- a/test/run
> > +++ b/test/run
> > @@ -67,6 +67,14 @@ run() {
> > test build/clang_tidy
> > teardown build
> >
> > + VALGRIND=0
> > + setup passt_in_ns
> > + test passt/ndp
> > + test passt/dhcp
> > + test perf/pasta_udp
> > + test passt_in_ns/shutdown
> > + teardown passt_in_ns
> > +
> > setup pasta
> > test pasta/ndp
> > test pasta/dhcp
>
> Ah, ok. Can we add that to the standard set of tests ASAP, please.
>
> > I get 21.6 gbps after this series, and 29.7 gbps before -- it's quite
> > significant.
>
> Drat.
>
> > And there's nothing strange in perf's output, really, the distribution
> > of overhead per functions is pretty much the same, but writing multiple
> > messages to the tap device just takes more cycles per message compared
> > to a single message.
>
> That's so weird. It should be basically an identical set of write()s,
> except that they happen in a batch, rather than a bit spread out. I
> guess it has to be some kind of cache locality thing. I wonder if the
> difference would go away or reverse if we had a way to submit multiple
> frames with a single syscall.
>
> > I'm a bit ashamed to propose this, but do you think about something
> > like:
>
> > if (c->mode == MODE_PASTA) { if (recvmmsg(ref.r.s, mmh_recv,
> > 1, 0, NULL) <= 0) return;
>
> > if (udp_mmh_splice_port(v6, mmh_recv)) { n =
> > recvmmsg(ref.r.s, mmh_recv + 1, UDP_MAX_FRAMES
> > - 1, 0, NULL); }
>
> > if (n > 0) n++; else n = 1; } else { n =
> > recvmmsg(ref.r.s, mmh_recv, UDP_MAX_FRAMES, 0,
> > NULL); if (n <= 0) return; }
>
> > ? Other than the inherent ugliness, it looks like a good
> > approximation to me.
>
> Hmm. Well, the first question is how much impact does going 1 message
> at a time have on the spliced throughput. If it's not too bad, then
> we could just always go one at a time for pasta, regardless of
> splicing. And we could even abstract that difference into the tap
> backend with a callback like tap_batch_size(c).
So, finally I had the chance to try this out.
First off, baseline with the patch adding the new tests I just sent,
and the series you posted:
=== perf/pasta_udp
> pasta: throughput and latency (local traffic)
Throughput in Gbps, latency in µs, one thread at 3.6 GHz, 4 streams
MTU: | 1500B | 4000B | 16384B | 65535B |
|--------|--------|--------|--------|
UDP throughput over IPv6: ns to host | 4.4 | 8.5 | 19.5 | 23.0 |
UDP RR latency over IPv6: ns to host | - | - | - | 27 |
|--------|--------|--------|--------|
UDP throughput over IPv4: ns to host | 4.3 | 8.8 | 18.5 | 24.4 |
UDP RR latency over IPv4: ns to host | - | - | - | 26 |
|--------|--------|--------|--------|
UDP throughput over IPv6: host to ns | - | - | - | 22.5 |
UDP RR latency over IPv6: host to ns | - | - | - | 30 |
|--------|--------|--------|--------|
UDP throughput over IPv4: host to ns | - | - | - | 24.5 |
UDP RR latency over IPv4: host to ns | - | - | - | 25 |
'--------'--------'--------'--------'
...passed.
> pasta: throughput and latency (traffic via tap)
Throughput in Gbps, latency in µs, one thread at 3.6 GHz, 4 streams
MTU: | 1500B | 4000B | 16384B | 65520B |
|--------|--------|--------|--------|
UDP throughput over IPv6: ns to host | 4.4 | 10.4 | 16.0 | 23.4 |
UDP RR latency over IPv6: ns to host | - | - | - | 27 |
|--------|--------|--------|--------|
UDP throughput over IPv4: ns to host | 5.2 | 10.8 | 16.0 | 24.0 |
UDP RR latency over IPv4: ns to host | - | - | - | 28 |
|--------|--------|--------|--------|
UDP throughput over IPv6: host to ns | - | - | - | 21.5 |
UDP RR latency over IPv6: host to ns | - | - | - | 29 |
|--------|--------|--------|--------|
UDP throughput over IPv4: host to ns | - | - | - | 26.3 |
UDP RR latency over IPv4: host to ns | - | - | - | 26 |
'--------'--------'--------'--------'
which seems to indicate the whole "splicing" thing is pretty much
useless, for UDP (except for that 16 KiB MTU case, but I wonder how
relevant that is).
If I set UDP_MAX_FRAMES to 1, with a quick workaround for the resulting
warning in udp_tap_send() (single frame to send, hence single message),
it gets somewhat weird:
=== perf/pasta_udp
> pasta: throughput and latency (local traffic)
Throughput in Gbps, latency in µs, one thread at 3.6 GHz, 4 streams
MTU: | 1500B | 4000B | 16384B | 65535B |
|--------|--------|--------|--------|
UDP throughput over IPv6: ns to host | 3.4 | 7.0 | 21.6 | 31.6 |
UDP RR latency over IPv6: ns to host | - | - | - | 30 |
|--------|--------|--------|--------|
UDP throughput over IPv4: ns to host | 3.8 | 7.0 | 22.0 | 32.4 |
UDP RR latency over IPv4: ns to host | - | - | - | 26 |
|--------|--------|--------|--------|
UDP throughput over IPv6: host to ns | - | - | - | 29.3 |
UDP RR latency over IPv6: host to ns | - | - | - | 31 |
|--------|--------|--------|--------|
UDP throughput over IPv4: host to ns | - | - | - | 33.8 |
UDP RR latency over IPv4: host to ns | - | - | - | 25 |
'--------'--------'--------'--------'
...passed.
> pasta: throughput and latency (traffic via tap)
Throughput in Gbps, latency in µs, one thread at 3.6 GHz, 4 streams
MTU: | 1500B | 4000B | 16384B | 65520B |
|--------|--------|--------|--------|
UDP throughput over IPv6: ns to host | 4.7 | 10.3 | 16.0 | 24.0 |
UDP RR latency over IPv6: ns to host | - | - | - | 27 |
|--------|--------|--------|--------|
UDP throughput over IPv4: ns to host | 5.6 | 11.4 | 16.0 | 24.0 |
UDP RR latency over IPv4: ns to host | - | - | - | 26 |
|--------|--------|--------|--------|
UDP throughput over IPv6: host to ns | - | - | - | 21.5 |
UDP RR latency over IPv6: host to ns | - | - | - | 29 |
|--------|--------|--------|--------|
UDP throughput over IPv4: host to ns | - | - | - | 28.7 |
UDP RR latency over IPv4: host to ns | - | - | - | 29 |
'--------'--------'--------'--------'
...except for the cases with low MTUs, throughput is significantly
higher if we read and send one message at a time on the "spliced" path.
Next, I would like to:
- bisect between 32 and 1 for UDP_MAX_FRAMES: maybe 32 affects data
locality too much, but some lower value would still be beneficial by
lowering syscall overhead
- try with sendmsg() instead of sendmmsg(), at this point. Looking at
the kernel, that doesn't seem to make a real difference.
About this series: should we just go ahead and apply it with
UDP_MAX_FRAMES set to 1 for the moment being? It's anyway better than
the existing situation.
--
Stefano
next prev parent reply other threads:[~2023-01-04 0:08 UTC|newest]
Thread overview: 28+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-12-05 8:14 [PATCH 0/8] Don't use additional sockets for receiving "spliced" UDP communications David Gibson
2022-12-05 8:14 ` [PATCH 1/8] udp: Move sending pasta tap frames to the end of udp_sock_handler() David Gibson
2022-12-05 8:14 ` [PATCH 2/8] udp: Split sending to passt tap interface into separate function David Gibson
2022-12-05 8:14 ` [PATCH 3/8] udp: Split receive from preparation and send in udp_sock_handler() David Gibson
2022-12-05 8:14 ` [PATCH 4/8] udp: Receive multiple datagrams at once on the pasta sock->tap path David Gibson
2022-12-13 22:48 ` Stefano Brivio
2022-12-14 1:42 ` David Gibson
2022-12-14 10:35 ` Stefano Brivio
2022-12-20 10:42 ` Stefano Brivio
2022-12-21 6:00 ` David Gibson
2022-12-22 2:37 ` Stefano Brivio
2023-01-04 0:08 ` Stefano Brivio [this message]
2023-01-04 4:53 ` David Gibson
2022-12-05 8:14 ` [PATCH 5/8] udp: Pre-populate msg_names with local address David Gibson
2022-12-13 22:48 ` Stefano Brivio
2022-12-14 1:22 ` David Gibson
2022-12-05 8:14 ` [PATCH 6/8] udp: Unify udp_sock_handler_splice() with udp_sock_handler() David Gibson
2022-12-13 22:48 ` Stefano Brivio
2022-12-14 1:19 ` David Gibson
2022-12-14 10:35 ` Stefano Brivio
2022-12-05 8:14 ` [PATCH 7/8] udp: Decide whether to "splice" per datagram rather than per socket David Gibson
2022-12-13 22:49 ` Stefano Brivio
2022-12-14 1:47 ` David Gibson
2022-12-14 10:35 ` Stefano Brivio
2022-12-15 0:33 ` David Gibson
2022-12-05 8:14 ` [PATCH 8/8] udp: Don't use separate sockets to listen for spliced packets David Gibson
2022-12-06 6:45 ` [PATCH 0/8] Don't use additional sockets for receiving "spliced" UDP communications Stefano Brivio
2022-12-06 6:46 ` Stefano Brivio
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=20230104010852.02e96a70@elisabeth \
--to=sbrivio@redhat.com \
--cc=david@gibson.dropbear.id.au \
--cc=passt-dev@passt.top \
/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.
Code repositories for project(s) associated with this public inbox
https://passt.top/passt
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 IMAP folder(s).