Performance issues with large timeshift depths

[torerikal]

We have been testing Shaka player with live streams of different, but rather large DVR windows. We notice a significantly increased CPU load with 1-4 hour window length. With 24 hours window length, Chrome becomes unresponsive in periods.

It looks like the CPU load increases proportionally with the DVR window length. Our testing is based on some private DASH streams, and unfortunately I can't share a manifest URL.

The 24 hours DVR window length will not be subject to real use cases. Still, we are concerned about the increased CPU load for live streams with 1 to 4 hours lengths, when attempted played on medium or low end hardware, and some special use cases with other concurrent CPU load. We also look into a workaround that might exacerbate the performance problem.

Some initial CPU profiling indicates that most CPU time is spent when updating the manifest. If I were going to guess, maybe the time codes for every segment in the timeline are re-computed every time the manifest is updated? If so, do you see a workaround or strategy that could be applied here?

[tdrews]

Thanks for the report.

If I were going to guess, maybe the time codes for every segment in the timeline are re-computed every time the manifest is updated?

Every segment specified in the new manifest is parsed into a SegmentReference, then after that, all those SegmentReferences are merged into the existing indexes. There are definitely some optimization opportunities, e.g., we could skip parsing segments that we know we have existing SegmentReferences for.

We'll discuss internally on what we can do here.

[raskri]

Why is this not a bug? As the stream is stuttering in Chrome and not in Firefox.

[joeyparrish]

If it's stuttering in Chrome and not Firefox, then one browser is executing the code faster or with less resource contention. Firefox being faster does not equal a bug in Shaka, in my opinion.

The merging code is semantically correct as far as I can tell, but we could/should make the library more efficient if we can. This is an enhancement, in my opinion.

The issue is currently tagged with "enhancement", which I believe is the best label for a potential performance improvement.

We have not investigated yet to determine what the cause is, but we'll try to get this fix into v2.0's final release. Since the code in question is completely new in v2 beta, it shouldn't affect v1.6.x, which is still the most stable release for now.

Sorry for the confusion over labels, and we'll keep you posted as we make progress.

[peyoh]

Hi Joe,
we also observed this in v.2
Seems like this one:

#288

[TheModMaker]

I just added a change to optimize the parser which increased performance in our tests. Can you try again from master and see if this fixes it?

[torerikal]

Thanks, this fix seems to have a big impact.

Without the fix, I have had CPU spikes with 100 % usage for a second or two. This is eliminated when applying the fix.

[peyoh]

Thanks,
now is better.