[Libwebsockets] browser flow-control

Per Bothner per at bothner.com
Sun Mar 12 20:58:46 CET 2017


On 03/12/2017 11:02 AM, Andy Green wrote:
> There's something not quite right about "bandwidth does not slow things down" and at the same time "the browser can't keep up".  Eventually, if it can't keep up, ACKs come too slowly to make space in the tcp window and POLLOUT isn't signalled until some ACKs come.
>
> It sounds like there's a buffer of a large size relative to the messages in the browser, that keeps the ACKs coming for a good while even while you can't keep up.

The problem isn't that messages get dropped, or we're not handling them correctly.
The problem is that the browser (at least Chrome and Firefox) never seem to get around to handling
keyboard and other user interface events.  For example if I run the 'yes' command
I can't ctrl-c to interrupt it.

> Apologies if the following doesn't apply... but...I guess if onMessage() did no work, you don't have any problem.  And getting data quickly is not necessarily a bad thing within reason.
>
> If so the problem is your onMessage() handling is expensive and the remote peer can spam triggering expensive work in the browser.
>
> How about decouple accepting data at onMessage() from the expensive work?  So eg if re-render or update of UI elements is the expensive bit, decouple re-render or update of the UI elements from logical handling of the incoming data.  Eat / process the data at onMessage() cheaply but then queue a rate-limited "expensive bit" on a timer and return from onMessage().

I really do want to slow down the sender if the browser can't keep up.
Yes, I can have onmessage just save the date in a queue and quickly return,
but if the onmessage calls come too quickly deferring the expensive display
update just makes things worse, as far as I can see.

>
>> I've implemented a builtin 'less'-like pager for ldomterm, and the plan
>> is to integrate
>> this into the flow-control: In auto-pager mode, automatic scroll
>> suspends display update
>> and enters paging mode, until the users scrolls forward to enable more
>> output.
>
> Sounds like they're two separate problems... if no ^S coming, browser should hopefully logically keep up even under endless spamming, even if it does not try update the UI fast enough to render every bit of the spam.  I use VNC on a tablet to a beefy machine for all my work the last few years, when VNC cant keep up with spew in a Gnome terminal window he acts well showing me a consistent snapshot of the spam (ie, no tearing in it itself) and then a clean view of how it ended.  This is very acceptable for the user.

This happens with DomTerm too, for example if I cat a very large file.
And that is acceptable.

But what if the file is infinite? I.e. it's printing in an endless loop?
I could tolerate this too, as long as the user can type ctrl-C, and that
is handled and gets sent to the looping process, and kills that process.
But the ctrl-C isn't getting through - I'm not sure why, but I suspect
it's just a scheduler in the browser that isn't designed to handle these situations.
-- 
	--Per Bothner
per at bothner.com   http://per.bothner.com/



More information about the Libwebsockets mailing list