httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stefan Eissing <stefan.eiss...@greenbytes.de>
Subject Re: HTTP/2 frame prioritization not honored
Date Sat, 28 Jan 2017 14:09:28 GMT
Kyriakis, I published a page using *your* graphics: https://icing.github.io/mod_h2/nimble.html
Please have a look. Anything you'd like to add or changed, just make a PR!

Thanks for the investigations and data!

Cheers,

Stefan

> Am 21.01.2017 um 11:50 schrieb Stefan Eissing <stefan.eissing@greenbytes.de>:
> 
> 
>> Am 20.01.2017 um 03:57 schrieb Kyriakos Zarifis <kyr.zarifis@gmail.com>:
>> 
>> * ... and by "here" I meant "here" 
> 
> :)
> 
>> On Thu, Jan 19, 2017 at 6:54 PM, Kyriakos Zarifis <kyr.zarifis@gmail.com> wrote:
>> Sounds great!
>> 
>> 
>> Very interested. I'd like to add a page over at https://icing.github.io/mod_h2/ about
it, so that people can easily grasp what the advantages are. For that, your numbers (do you
have screenshots of browser timelines maybe?) would be very welcome. Also that someone besides
the module author has measured it adds credibility. :-)
>> 
>> If you write yourself somewhere about it, I am happy to link that.
>> 
>> Since anything I write would be incomplete without your description of what caused
it and how you resolved it, I put together a WIP write up here, with screenshots / link to
logs*. Feel free to use it as you want or let me know if you'd like more details; I'm happy
to help write the complete story for a page on https://icing.github.io/mod_h2/ , which is
probably the most reasonable place to gather the relevant bits.
>> 
>> Cheers
>> 
>> 
>> 
>> * I rerun the tests to capture timeline screenshots, so the server-logs don't exactly
correspond to those screenshots, but the behaviors were the same
>> * Note that the delays in the timeline pictures are worse than those seen on server
logs, which have been more helpful for understanding application-layer behavior. I think what's
causing this is a bloated output buffer in the case where the server aggressively writes low-prio
data (verified this by monitoring the buffer size which keeps increasing during the test)
> 
> Certainly an area to improve upon. mod_http2 is still writing so much into the socket
that responsiveness suffers. This gives the best throughput performance, though. I know that
the h2o server guys also experimented with interrogating TCP windows to prevent bloat. Have
to look at that again.
> 
> Cheers,
> 
> Stefan Eissing
> 
> <green/>bytes GmbH
> Hafenstrasse 16
> 48155 Münster
> www.greenbytes.de
> 

Stefan Eissing

<green/>bytes GmbH
Hafenstrasse 16
48155 Münster
www.greenbytes.de


Mime
View raw message