MIME-Version: 1.0
References: <20230221110344.82818-1-kerneljasonxing@gmail.com>
 <48429c16fdaee59867df5ef487e73d4b1bf099af.camel@redhat.com>
 <CAL+tcoD8PzL4khHq44z27qSHHGkcC4YUa91E3h+ki7O0u3SshQ@mail.gmail.com>
 <aaf3d11ea5b247ab03d117dadae682fe2180d38a.camel@redhat.com> <CAL+tcoBZFFwOnUqzcDtSsNyfPgHENAOv0bPcvncxuMPwCn40+Q@mail.gmail.com>
In-Reply-To: <CAL+tcoBZFFwOnUqzcDtSsNyfPgHENAOv0bPcvncxuMPwCn40+Q@mail.gmail.com>
From:   Jason Xing <kerneljasonxing@gmail.com>
Date:   Wed, 22 Feb 2023 11:47:06 +0800
Message-ID: <CAL+tcoBGFkXea-GyzbO41Ve8_wUF3PT=YF43TxuzgM+adVa8gw@mail.gmail.com>
Subject: Re: [PATCH net] udp: fix memory schedule error
To:     Paolo Abeni <pabeni@redhat.com>
Cc:     willemdebruijn.kernel@gmail.com, davem@davemloft.net,
        dsahern@kernel.org, edumazet@google.com, kuba@kernel.org,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
        bpf@vger.kernel.org, Jason Xing <kernelxing@tencent.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Feb 21, 2023 at 11:46 PM Jason Xing <kerneljasonxing@gmail.com> wrote:
>
> On Tue, Feb 21, 2023 at 10:46 PM Paolo Abeni <pabeni@redhat.com> wrote:
> >
> > On Tue, 2023-02-21 at 21:39 +0800, Jason Xing wrote:
> > > On Tue, Feb 21, 2023 at 8:27 PM Paolo Abeni <pabeni@redhat.com> wrote:
> > > >
> > > > On Tue, 2023-02-21 at 19:03 +0800, Jason Xing wrote:
> > > > > From: Jason Xing <kernelxing@tencent.com>
> > > > >
> > > > > Quoting from the commit 7c80b038d23e ("net: fix sk_wmem_schedule()
> > > > > and sk_rmem_schedule() errors"):
> > > > >
> > > > > "If sk->sk_forward_alloc is 150000, and we need to schedule 150001 bytes,
> > > > > we want to allocate 1 byte more (rounded up to one page),
> > > > > instead of 150001"
> > > >
> > > > I'm wondering if this would cause measurable (even small) performance
> > > > regression? Specifically under high packet rate, with BH and user-space
> > > > processing happening on different CPUs.
> > > >
> > > > Could you please provide the relevant performance figures?
> > >
> > > Sure, I've done some basic tests on my machine as below.
> > >
> > > Environment: 16 cpus, 60G memory
> > > Server: run "iperf3 -s -p [port]" command and start 500 processes.
> > > Client: run "iperf3 -u -c 127.0.0.1 -p [port]" command and start 500 processes.
> >
> > Just for the records, with the above command each process will send
> > pkts at 1mbs - not very relevant performance wise.
> >
> > Instead you could do:
> >
>
> > taskset 0x2 iperf -s &
> > iperf -u -c 127.0.0.1 -b 0 -l 64
> >
>
> Thanks for your guidance.
>
> Here're some numbers according to what you suggested, which I tested
> several times.
> ----------|IFACE   rxpck/s   txpck/s    rxkB/s    txkB/s
> Before: lo 411073.41 411073.41  36932.38  36932.38
> After:   lo 410308.73 410308.73  36863.81  36863.81
>
> Above is one of many results which does not mean that the original
> code absolutely outperforms.
> The output is not that constant and stable, I think.

Today, I ran the same test on other servers, it looks the same as
above. Those results fluctuate within ~2%.

Oh, one more thing I forgot to say is the output of iperf itself which
doesn't show any difference.
Before: Bitrate is 211 - 212 Mbits/sec
After: Bitrate is 211 - 212 Mbits/sec
So this result is relatively constant especially if we keep running
the test over 2 minutes.

Jason

>
> Please help me review those numbers.
>
> >
> > > In theory, I have no clue about why it could cause some regression?
> > > Maybe the memory allocation is not that enough compared to the
> > > original code?
> >
> > As Eric noted, for UDP traffic, due to the expected average packet
> > size, sk_forward_alloc is touched quite frequently, both with and
> > without this patch, so there is little chance it will have any
> > performance impact.
>
> Well, I see.
>
> Thanks,
> Jason
>
> >
> > Cheers,
> >
> > Paolo
> >