netdev - Re: [PATCH v2] tcp: adjust rcvbuff according copied rate of user space

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <CADVnQynuFfW5DAeHM1LwKh0YgG3A27Zn2H6MGsj_14WdXCMZsA@mail.gmail.com>
Date:   Thu, 11 Aug 2022 09:26:59 -0400
From:   Neal Cardwell <ncardwell@...gle.com>
To:     Yonglong Li <liyonglong@...natelecom.cn>
Cc:     netdev@...r.kernel.org, davem@...emloft.net, edumazet@...gle.com,
        ycheng@...gle.com, dsahern@...nel.org, kuba@...nel.org,
        pabeni@...hat.com
Subject: Re: [PATCH v2] tcp: adjust rcvbuff according copied rate of user space

On Thu, Aug 11, 2022 at 4:15 AM Yonglong Li <liyonglong@...natelecom.cn> wrote:
>
>
>
> On 8/10/2022 8:43 PM, Neal Cardwell wrote:
> > On Wed, Aug 10, 2022 at 3:49 AM Yonglong Li <liyonglong@...natelecom.cn> wrote:
> >>
> >> every time data is copied to user space tcp_rcv_space_adjust is called.
> >> current It adjust rcvbuff by the length of data copied to user space.
> >> If the interval of user space copy data from socket is not stable, the
> >> length of data copied to user space will not exactly show the speed of
> >> copying data from rcvbuff.
> >> so in tcp_rcv_space_adjust it is more reasonable to adjust rcvbuff by
> >> copied rate (length of copied data/interval)instead of copied data len
> >>
> >> I tested this patch in simulation environment by Mininet:
> >> with 80~120ms RTT / 1% loss link, 100 runs
> >> of (netperf -t TCP_STREAM -l 5), and got an average throughput
> >> of 17715 Kbit instead of 17703 Kbit.
> >> with 80~120ms RTT without loss link, 100 runs of (netperf -t
> >> TCP_STREAM -l 5), and got an average throughput of 18272 Kbit
> >> instead of 18248 Kbit.
> >
> > So with 1% emulated loss that's a 0.06% throughput improvement and
> > without emulated loss that's a 0.13% improvement. That sounds like it
> > may well be statistical noise, particularly given that we would expect
> > the steady-state impact of this change to be negligible.
> >
> Hi neal,
>
> Thank you for your feedback.
> I don't think the improvement is statistical noise. Because I can get small
> improvement after patch every time I test.

Interesting. To help us all understand the dynamics, can you please
share a sender-side tcpdump binary .pcap trace of the emulated tests
without loss, with:

(a) one baseline pcap of a test without the patch, and

(b) one experimental pcap of a test with the patch showing the roughly
0.13% throughput improvement.

It will be interesting to compare the receive window and transmit
behavior in both cases.

thanks,
neal