[foms] What data is needed for adaptive stream switching?

Mark Watson watsonm at netflix.com
Tue Nov 30 14:47:03 PST 2010

On Nov 29, 2010, at 8:05 AM, Frank Galligan wrote:

On Tue, Nov 23, 2010 at 4:24 PM, Chris Pearce <chris at pearce.org.nz<mailto:chris at pearce.org.nz>> wrote:
Thanks for explaining Mark. Much appreciated.

On 24/11/2010 6:51 a.m., Mark Watson wrote:
> This is where there is scope for experimentation. What I think would be great is to define an API which can indicate these decision points, provide the two data sets (past incoming bandwidth) and (future bitrate of each stream) at some sufficient level of generality and indicate back the decision. Then we can experiment with more and less complex input data and more and less complex decision algorithms.

So in terms of what changes to browsers we'd need to make to start
experimenting, we'd need to resurrect the @bufferedBytes attribute, add
a @currentOffset attribute, and add some way for JS to access the
RAP/keyframe index?
I'm not sure what exposing the RAP to the JS for each stream buys you over just having a simple switch stream api. Let me back up a bit.

At a high level I think we want the media pipeline to expose information about the current presentation to the player through a script interface. Then the player can make decisions about what stream the media pipeline should be rendering. The decisions made by the player are it's best guess so the media pipeline will not stall for any reason. We want to do this because trying to define an algorithm that all default players must implement is going to be extremely hard. This is my view from reading over the posts on this list.

So to me I don't think that a JS player really cares that the media pipeline switches to stream N at byte offset X (of course the media pipeline does). All the player cares about is that the media pipeline switch to stream N as soon as possible. I.E. The player decides that the CPU load is too great and wants to switch to a lower resolution as soon as possible seamlessly. Or the player decides that the bandwidth looks high enough that it would like to switch to a higher bandwidth stream as soon as possible seamlessly.

I think having an API that is SwitchTo(index, discontinuous); should be good enough for now. (I added a discontinuous parameter in case the player wanted to switch and didn't care about it being seamless.) Are there any reasons we need to expose RAPs to the player?

The problem if you make the information exposed to the JS layer too simple is that you end up with essentially only one adaptation algorithm (compare incoming bandwidth to stream rates) and then there is no scope for experimentation.

On the other hand, I agree that it could easily be made too complex.

So the question is how "precise" do we expect the adaptation algorithms at the JS level to be ? Exposing the switch points is of interest if you expect the JS algorithm to be making quite precise decisions based on the current amount of received data, remaining distance to the next switch point and precise VBR profile of the streams going forward.

I can think of a corner case, but I don't think the added complexity to the interface and to the player developer is worth it. The case I'm thinking of is the player wants to switch to a lower bandwidth stream but stream N which the player wants to switch to doesn't have a RAP for 10 seconds while stream N-1, which has bandwidth < stream N, has a RAP 2 seconds out. If people felt strongly about handling this case we could expose an API like SwitchDown(attribute); Then there would also be an API call that can control the heuristics if the media pipeline had to switch to another stream with a lower value attribute than the desired stream.

I'm not sure it's a corner case. Even if I have switch points every 2s if my bandwidth drops by 50% then it could be as much as 4s to receive enough data to get to the next switch point. I probably need to account for this and I may choose an even lower stream rate than otherwise to avoid a stall.

Maybe we should add the @bufferedBytes data into
@buffered, so you can easily map buffered time ranges to byte ranges? I
guess these would have to be per-stream if we're playing multiple
independent streams.

Or would you prefer an explicit download bandwidth and a per stream
bitrate measure, calculated by the browser, over a specified time window?

In terms of an API which can indicate decision points, maybe an event
which fires when playback enters a new chunk? Or fires when the download
of a chunk finishes? Be aware we can't guarantee that any DOM events
which we fire will arrive in a particularly timely manner, there could
be any number of other things going on with the event queue.
This is another reason that I think a higher level API for the player is better.

Is there a direct mapping between a keyframe index and RAP points? What
about audio streams in containers which don't index audio? Particularly
in the case where we're playing multiple independent streams.
For WebM I wrote a tool that would create an index file for an audio stream.


Chris P.

foms mailing list
foms at lists.annodex.net<mailto:foms at lists.annodex.net>


-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.annodex.net/cgi-bin/mailman/private/foms/attachments/20101130/3bf1da63/attachment.htm 

More information about the foms mailing list