You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@qpid.apache.org by Michael Goulish <mg...@redhat.com> on 2021/02/16 18:53:04 UTC
Router Throughput vs Buffer Size -- 11 data points (2nd try)
I keep forgetting that I can't include images.
<https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0>
Here it is.
<https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0>
I'm afraid there's not much of any knee in this curve.
By the way, CPU was almost unaffected.
Actually it improved slightly, from 208% at 512 byte buffer,
to 195 for buf=2048 to 191 for buf=3072 and above.
Re: Router Throughput vs Buffer Size -- 11 data points (2nd try)
Posted by Chuck Rolke <cr...@redhat.com>.
Http2 adaptor uses 16k (or so) -byte buffers between the adaptor and proton
raw connection. The TCP adaptor could probably use the http2 buffering scheme
and amortize the cost of doing a buffer copy between the 16k byte buffers and
the normal dispatch buffers by doing fewer scheduling turns and still be way
ahead.
Vectored IO would be an improvement to any buffer size.
Dispatch's choice of 512-byte buffers is probably due for a revisit. Simply
changing that to 2k, and adjusting any number of assumptions based on the
512-byte size, could provide a simple and effective boost to busy router
performance.
----- Original Message -----
> From: "Ted Ross" <tr...@apache.org>
> To: users@qpid.apache.org
> Sent: Tuesday, February 16, 2021 2:26:03 PM
> Subject: Re: Router Throughput vs Buffer Size -- 11 data points (2nd try)
>
> Thanks for this Mick,
>
> The original context of the request for this graph was the AMQP
> large-message testing. Was this test run using iperf over the TCP
> adaptor? If so, it's possible that the buffering between the adaptor and
> Proton is obscuring the results.
>
> I suspect that in the absence of vectored IO, the optimal Proton
> raw-buffer-size is larger than the optimal internal-buffer-size. This is a
> conjecture not based on science.
>
> -Ted
>
> On Tue, Feb 16, 2021 at 1:53 PM Michael Goulish <mg...@redhat.com> wrote:
>
> > I keep forgetting that I can't include images.
> >
> > <https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0>
> > Here it is.
> > <https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0>
> >
> > I'm afraid there's not much of any knee in this curve.
> >
> > By the way, CPU was almost unaffected.
> > Actually it improved slightly, from 208% at 512 byte buffer,
> > to 195 for buf=2048 to 191 for buf=3072 and above.
> >
>
---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
For additional commands, e-mail: users-help@qpid.apache.org
Re: Router Throughput vs Buffer Size -- 11 data points (2nd try)
Posted by Michael Goulish <mg...@redhat.com>.
Yes, this was with iperf3 -- I will see what I can do with AMQP.
On Tue, Feb 16, 2021 at 2:32 PM Ted Ross <tr...@apache.org> wrote:
> Thanks for this Mick,
>
> The original context of the request for this graph was the AMQP
> large-message testing. Was this test run using iperf over the TCP
> adaptor? If so, it's possible that the buffering between the adaptor and
> Proton is obscuring the results.
>
> I suspect that in the absence of vectored IO, the optimal Proton
> raw-buffer-size is larger than the optimal internal-buffer-size. This is a
> conjecture not based on science.
>
> -Ted
>
> On Tue, Feb 16, 2021 at 1:53 PM Michael Goulish <mg...@redhat.com>
> wrote:
>
> > I keep forgetting that I can't include images.
> >
> > <https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0
> >
> > Here it is.
> > <https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0
> >
> >
> > I'm afraid there's not much of any knee in this curve.
> >
> > By the way, CPU was almost unaffected.
> > Actually it improved slightly, from 208% at 512 byte buffer,
> > to 195 for buf=2048 to 191 for buf=3072 and above.
> >
>
Re: Router Throughput vs Buffer Size -- 11 data points (2nd try)
Posted by Ted Ross <tr...@apache.org>.
Thanks for this Mick,
The original context of the request for this graph was the AMQP
large-message testing. Was this test run using iperf over the TCP
adaptor? If so, it's possible that the buffering between the adaptor and
Proton is obscuring the results.
I suspect that in the absence of vectored IO, the optimal Proton
raw-buffer-size is larger than the optimal internal-buffer-size. This is a
conjecture not based on science.
-Ted
On Tue, Feb 16, 2021 at 1:53 PM Michael Goulish <mg...@redhat.com> wrote:
> I keep forgetting that I can't include images.
>
> <https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0>
> Here it is.
> <https://www.dropbox.com/s/b9sbe4cbh3mddo4/buffer_vs_throughput.jpg?dl=0>
>
> I'm afraid there's not much of any knee in this curve.
>
> By the way, CPU was almost unaffected.
> Actually it improved slightly, from 208% at 512 byte buffer,
> to 195 for buf=2048 to 191 for buf=3072 and above.
>