On Sun, Jun 17, 2018 at 12:30 PM Brian Wolff <bawolff(a)gmail.com> wrote: :D Using the same CPU thread count, typical files take about 3-4 times longer with the configuration we've got ready. This doesn't seem to be a major problem at low resolutions, but higher resolutions may need to have the limits raised for very long videos like conference talks. Memory usage may be higher but I've not specifically noticed a major difference between VP8 and VP9 there. This could be improved significantly by updating to libvpx 1.7 and a matching version of ffmpeg that supports macroblock row multithreading: this means that we'll be able to use more than 4 threads for HD videos, which should allow using all cores on a 20-core/40-thread machine if it's not loaded up with other files. The necessary libraries are released; we just need to backport the newer packages to Debian 9 I believe. Don't yet know whether this will be an easy or hard task. Now that you mention it I am concerned about the time limit on very long videos, so I'll take another look at the packaging backport and see if we can knock that out quickly. :) -- brion

Current state on this: * still hoping to deploy the libvpx+ffmpeg backport first so we start with best performance; Moritz made a start on libvpx but we still have to resolve ffmpeg (possibly by patching 3.2 instead of updating all the way to 3.4)

Awesome sauce. Thanks Moritz! -- brion On Fri, Jun 29, 2018 at 7:39 AM Moritz Muehlenhoff < mmuhlenhoff(a)wikimedia.org&gt; wrote:

Oh and one more thing! For the VP9 configuration I'll be enabling 1440p and 2160p ("4K") resolutions, which people can manually bump up to when watching videos with a suitable 4K source on a high-res screen. They use higher data rates, but only a small fraction of input files are 4K so should not significantly increase disk space projections for now. These can take a long time to compress, so if we find it's problematic we'll turn them back off until the jobs can be split into tiny chunks (future work planned!), but it works in my testing and shouldn't clog the servers now that we have more available. (Note that the ogv.js player shim for Safari will not handle greater-than-HD resolutions fast enough for playback, even on a fast Mac or iPad; for best results for 4K playback use Firefox, Chrome, or a Chromium-based browser.) -- brion On Thu, Jul 26, 2018 at 5:39 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; wrote: _______________________________________________ Multimedia mailing list Multimedia(a)lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/multimedia

Oh and one more thing! For the VP9 configuration I'll be enabling 1440p and 2160p ("4K") resolutions, which people can manually bump up to when watching videos with a suitable 4K source on a high-res screen. They use higher data rates, but only a small fraction of input files are 4K so should not significantly increase disk space projections for now. These can take a long time to compress, so if we find it's problematic we'll turn them back off until the jobs can be split into tiny chunks (future work planned!), but it works in my testing and shouldn't clog the servers now that we have more available. (Note that the ogv.js player shim for Safari will not handle greater-than-HD resolutions fast enough for playback, even on a fast Mac or iPad; for best results for 4K playback use Firefox, Chrome, or a Chromium-based browser.) -- brion On Thu, Jul 26, 2018 at 5:39 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; wrote: > Ok, after some delay for re-tweaking the encoding settings for higher > quality when needed, and pulling in some other improvements to the config > system, all related updates to TimedMediaHandler have been merged. :D > > If all goes well with the general deployments in the next few days, > expect the beginning of VP9 rollout starting next week. > > Changes since the earlier announcement: > * the new row-multithreading will be available, which allows higher > threading usage at all resolutions; encoding times will be more like 1.5-2x > slower instead of 3-4x slower. > * switch to constrained quality with a larger max bitrate: many files > will become significantly smaller in their VP9 versions, but some will > actually increase in exchange for a huge increase in quality -- this is > mostly 60fps high-rate files, and those with lots of motion and detail that > didn't compress well at the default low data rates. > > -- brion > > On Fri, Jun 29, 2018 at 9:46 AM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; > wrote: > >> Awesome sauce. Thanks Moritz! >> >> -- brion >> >> On Fri, Jun 29, 2018 at 7:39 AM Moritz Muehlenhoff < >> mmuhlenhoff(a)wikimedia.org&gt; wrote: >> >>> Hi all, >>> >>> On Thu, Jun 28, 2018 at 01:54:18PM -0700, Brion Vibber wrote: >>> > Current state on this: >>> > >>> > * still hoping to deploy the libvpx+ffmpeg backport first so we start >>> with >>> > best performance; Moritz made a start on libvpx but we still have to >>> > resolve ffmpeg (possibly by patching 3.2 instead of updating all the >>> way to >>> > 3.4) >>> >>> I've completed this today. We now have a separate repository component >>> for stretch-wikimedia (named component/vp9) which includes ffmpeg 3.2.10 >>> (thus allowing us to follow the ffmpeg security updates released in >>> Debian >>> with a local rebuild) with backported row-mt support and linked against >>> libvpx 1.7.0. >>> >>> I tested re-encoding >>> >>> https://commons.wikimedia.org/wiki/File:Wall_of_Death_-_Pitts_Todeswand_201… >>> (which is a nice fast-paced test file) from VP8 to VP9, which results in >>> a size reduction from 48M to 31M. >>> >>> When using eight CPU cores on one of our video scaler servers, enabling >>> row-mt >>> gives a significant performance boost; encoding time went down from >>> 5:31 mins >>> to 3:36 mins. >>> >>> All the details can be found at >>> https://phabricator.wikimedia.org/T190333#4324995 >>> >>> Cheers, >>> Moritz >>> >>> _______________________________________________ >>> Wikitech-l mailing list >>> Wikitech-l(a)lists.wikimedia.org >>> https://lists.wikimedia.org/mailman/listinfo/wikitech-l >> >>

Batch process is continuing... over 2200 source videos have been compressed to VP9 so far, of the 120k+ total in the system. Seeing big improvements in overall compression, though it's hard to tell how representative the subset of conversions done so far is. I'll post occasional updates to the transcode report charts at: https://www.mediawiki.org/wiki/Extension:TimedMediaHandler/VP9_transition/r… In the histograms you can see that not only is the average bitrate down for VP9 vs VP8, but there's a larger "spread" between low and high bandwidth in the VP9 versions thanks to the constrained-quality configuration vs the old fixed bitrate target. This allows files that compress well to use fewer bits, while those that have high frame rates or lots of detail/motion use more bits to encode higher quality than they did before. It will take at least a few weeks to recompress everything, and we may or may not want to increase the number of simultaneous jobs (so far the servers are running beautifully, but are a bit under-loaded). -- brion On Mon, Jul 30, 2018 at 5:51 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; wrote: > Configuration has been updated and VP9 mode swapped in. Newly uploaded > videos should start encoding as VP9 starting now. > > I'll start the backfill batch process tonight or tomorrow, and will > likely stop and restart it a few times over the coming weeks if anything > needs adjustment. Please file tasks in phabricator and/or give me a ping on > IRC if any issues come up with the new conversions, or with old files! > (Existing VP8 files will be left as-is for now until we're sure > everything's up to snuff.) > > Big thanks to everybody who's helped prepping this, with libvpx and > ffmpeg deployments, with the patches and review, and with final deployment > which always takes longer than you expect. :) This'll be a nice improvement > to our video output for now, and lays a lot of groundwork for next steps. > > -- brion > > > On Thu, Jul 26, 2018 at 5:47 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; > wrote: > >> Oh and one more thing! >> >> For the VP9 configuration I'll be enabling 1440p and 2160p ("4K") >> resolutions, which people can manually bump up to when watching videos with >> a suitable 4K source on a high-res screen. They use higher data rates, but >> only a small fraction of input files are 4K so should not significantly >> increase disk space projections for now. >> >> These can take a long time to compress, so if we find it's problematic >> we'll turn them back off until the jobs can be split into tiny chunks >> (future work planned!), but it works in my testing and shouldn't clog the >> servers now that we have more available. >> >> (Note that the ogv.js player shim for Safari will not handle >> greater-than-HD resolutions fast enough for playback, even on a fast Mac or >> iPad; for best results for 4K playback use Firefox, Chrome, or a >> Chromium-based browser.) >> >> -- brion >> >> On Thu, Jul 26, 2018 at 5:39 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; >> wrote: >> >>> Ok, after some delay for re-tweaking the encoding settings for higher >>> quality when needed, and pulling in some other improvements to the config >>> system, all related updates to TimedMediaHandler have been merged. :D >>> >>> If all goes well with the general deployments in the next few days, >>> expect the beginning of VP9 rollout starting next week. >>> >>> Changes since the earlier announcement: >>> * the new row-multithreading will be available, which allows higher >>> threading usage at all resolutions; encoding times will be more like 1.5-2x >>> slower instead of 3-4x slower. >>> * switch to constrained quality with a larger max bitrate: many files >>> will become significantly smaller in their VP9 versions, but some will >>> actually increase in exchange for a huge increase in quality -- this is >>> mostly 60fps high-rate files, and those with lots of motion and detail that >>> didn't compress well at the default low data rates. >>> >>> -- brion >>> >>> On Fri, Jun 29, 2018 at 9:46 AM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; >>> wrote: >>> >>>> Awesome sauce. Thanks Moritz! >>>> >>>> -- brion >>>> >>>> On Fri, Jun 29, 2018 at 7:39 AM Moritz Muehlenhoff < >>>> mmuhlenhoff(a)wikimedia.org&gt; wrote: >>>> >>>>> Hi all, >>>>> >>>>> On Thu, Jun 28, 2018 at 01:54:18PM -0700, Brion Vibber wrote: >>>>> > Current state on this: >>>>> > >>>>> > * still hoping to deploy the libvpx+ffmpeg backport first so we >>>>> start with >>>>> > best performance; Moritz made a start on libvpx but we still have to >>>>> > resolve ffmpeg (possibly by patching 3.2 instead of updating all >>>>> the way to >>>>> > 3.4) >>>>> >>>>> I've completed this today. We now have a separate repository component >>>>> for stretch-wikimedia (named component/vp9) which includes ffmpeg >>>>> 3.2.10 >>>>> (thus allowing us to follow the ffmpeg security updates released in >>>>> Debian >>>>> with a local rebuild) with backported row-mt support and linked >>>>> against >>>>> libvpx 1.7.0. >>>>> >>>>> I tested re-encoding >>>>> >>>>> https://commons.wikimedia.org/wiki/File:Wall_of_Death_-_Pitts_Todeswand_201… >>>>> (which is a nice fast-paced test file) from VP8 to VP9, which results >>>>> in >>>>> a size reduction from 48M to 31M. >>>>> >>>>> When using eight CPU cores on one of our video scaler servers, >>>>> enabling row-mt >>>>> gives a significant performance boost; encoding time went down from >>>>> 5:31 mins >>>>> to 3:36 mins. >>>>> >>>>> All the details can be found at >>>>> https://phabricator.wikimedia.org/T190333#4324995 >>>>> >>>>> Cheers, >>>>> Moritz >>>>> >>>>> _______________________________________________ >>>>> Wikitech-l mailing list >>>>> Wikitech-l(a)lists.wikimedia.org >>>>> https://lists.wikimedia.org/mailman/listinfo/wikitech-l >>>> >>>>

These conversions have been running in the background, at a moderate load to ensure we don't slow down newer uploads much, and are up to the "P"s alphabetically so we're on schedule. :) I've temporarily stopped the batch process pending the upcoming datacenter switchover test, and will continue once it's switched. -- brion On Wed, Aug 8, 2018 at 2:06 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; wrote: > Quick update... no major problems noticed so far. One configuration > change I made was to re-enable VP8 transcodes for new files, since > disabling them completely lead to the existing ones not being used. Later > in the transition, or afterwards, we'll clean up the now-unused VP8 > transcodes. > > If I'm reading numbers in > https://www.mediawiki.org/wiki/Extension:TimedMediaHandler/VP9_ > transition/reports right, almost 5% of files by count or 12% by duration > have been converted so far. > > That leads to a remaining duration somewhere between 9 weeks (assuming 9 > days did 12% of stuff) and 24 weeks (9 days did 5% of stuff), depending on > actual factors like whether the remaining files are the same mix of > resolutions, and how we adjust the load balance on the job queue. > > -- brion > > On Fri, Aug 3, 2018 at 11:22 AM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; > wrote: > >> Batch process is continuing... over 2200 source videos have been >> compressed to VP9 so far, of the 120k+ total in the system. >> >> Seeing big improvements in overall compression, though it's hard to tell >> how representative the subset of conversions done so far is. I'll post >> occasional updates to the transcode report charts at: >> >> https://www.mediawiki.org/wiki/Extension:TimedMediaHandler/VP9_transition/r… >> >> In the histograms you can see that not only is the average bitrate down >> for VP9 vs VP8, but there's a larger "spread" between low and high >> bandwidth in the VP9 versions thanks to the constrained-quality >> configuration vs the old fixed bitrate target. This allows files that >> compress well to use fewer bits, while those that have high frame rates or >> lots of detail/motion use more bits to encode higher quality than they did >> before. >> >> It will take at least a few weeks to recompress everything, and we may >> or may not want to increase the number of simultaneous jobs (so far the >> servers are running beautifully, but are a bit under-loaded). >> >> -- brion >> >> >> On Mon, Jul 30, 2018 at 5:51 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; >> wrote: >> >>> Configuration has been updated and VP9 mode swapped in. Newly uploaded >>> videos should start encoding as VP9 starting now. >>> >>> I'll start the backfill batch process tonight or tomorrow, and will >>> likely stop and restart it a few times over the coming weeks if anything >>> needs adjustment. Please file tasks in phabricator and/or give me a ping on >>> IRC if any issues come up with the new conversions, or with old files! >>> (Existing VP8 files will be left as-is for now until we're sure >>> everything's up to snuff.) >>> >>> Big thanks to everybody who's helped prepping this, with libvpx and >>> ffmpeg deployments, with the patches and review, and with final deployment >>> which always takes longer than you expect. :) This'll be a nice improvement >>> to our video output for now, and lays a lot of groundwork for next steps. >>> >>> -- brion >>> >>> >>> On Thu, Jul 26, 2018 at 5:47 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; >>> wrote: >>> >>>> Oh and one more thing! >>>> >>>> For the VP9 configuration I'll be enabling 1440p and 2160p ("4K") >>>> resolutions, which people can manually bump up to when watching videos with >>>> a suitable 4K source on a high-res screen. They use higher data rates, but >>>> only a small fraction of input files are 4K so should not significantly >>>> increase disk space projections for now. >>>> >>>> These can take a long time to compress, so if we find it's problematic >>>> we'll turn them back off until the jobs can be split into tiny chunks >>>> (future work planned!), but it works in my testing and shouldn't clog the >>>> servers now that we have more available. >>>> >>>> (Note that the ogv.js player shim for Safari will not handle >>>> greater-than-HD resolutions fast enough for playback, even on a fast Mac or >>>> iPad; for best results for 4K playback use Firefox, Chrome, or a >>>> Chromium-based browser.) >>>> >>>> -- brion >>>> >>>> On Thu, Jul 26, 2018 at 5:39 PM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; >>>> wrote: >>>> >>>>> Ok, after some delay for re-tweaking the encoding settings for higher >>>>> quality when needed, and pulling in some other improvements to the config >>>>> system, all related updates to TimedMediaHandler have been merged. :D >>>>> >>>>> If all goes well with the general deployments in the next few days, >>>>> expect the beginning of VP9 rollout starting next week. >>>>> >>>>> Changes since the earlier announcement: >>>>> * the new row-multithreading will be available, which allows higher >>>>> threading usage at all resolutions; encoding times will be more like 1.5-2x >>>>> slower instead of 3-4x slower. >>>>> * switch to constrained quality with a larger max bitrate: many files >>>>> will become significantly smaller in their VP9 versions, but some will >>>>> actually increase in exchange for a huge increase in quality -- this is >>>>> mostly 60fps high-rate files, and those with lots of motion and detail that >>>>> didn't compress well at the default low data rates. >>>>> >>>>> -- brion >>>>> >>>>> On Fri, Jun 29, 2018 at 9:46 AM Brion Vibber &lt;bvibber(a)wikimedia.org&gt; >>>>> wrote: >>>>> >>>>>> Awesome sauce. Thanks Moritz! >>>>>> >>>>>> -- brion >>>>>> >>>>>> On Fri, Jun 29, 2018 at 7:39 AM Moritz Muehlenhoff < >>>>>> mmuhlenhoff(a)wikimedia.org&gt; wrote: >>>>>> >>>>>>> Hi all, >>>>>>> >>>>>>> On Thu, Jun 28, 2018 at 01:54:18PM -0700, Brion Vibber wrote: >>>>>>> > Current state on this: >>>>>>> > >>>>>>> > * still hoping to deploy the libvpx+ffmpeg backport first so we >>>>>>> start with >>>>>>> > best performance; Moritz made a start on libvpx but we still have >>>>>>> to >>>>>>> > resolve ffmpeg (possibly by patching 3.2 instead of updating all >>>>>>> the way to >>>>>>> > 3.4) >>>>>>> >>>>>>> I've completed this today. We now have a separate repository >>>>>>> component >>>>>>> for stretch-wikimedia (named component/vp9) which includes ffmpeg >>>>>>> 3.2.10 >>>>>>> (thus allowing us to follow the ffmpeg security updates released in >>>>>>> Debian >>>>>>> with a local rebuild) with backported row-mt support and linked >>>>>>> against >>>>>>> libvpx 1.7.0. >>>>>>> >>>>>>> I tested re-encoding >>>>>>> >>>>>>> https://commons.wikimedia.org/wiki/File:Wall_of_Death_-_Pitts_Todeswand_201… >>>>>>> (which is a nice fast-paced test file) from VP8 to VP9, which >>>>>>> results in >>>>>>> a size reduction from 48M to 31M. >>>>>>> >>>>>>> When using eight CPU cores on one of our video scaler servers, >>>>>>> enabling row-mt >>>>>>> gives a significant performance boost; encoding time went down from >>>>>>> 5:31 mins >>>>>>> to 3:36 mins. >>>>>>> >>>>>>> All the details can be found at >>>>>>> https://phabricator.wikimedia.org/T190333#4324995 >>>>>>> >>>>>>> Cheers, >>>>>>> Moritz >>>>>>> >>>>>>> _______________________________________________ >>>>>>> Wikitech-l mailing list >>>>>>> Wikitech-l(a)lists.wikimedia.org >>>>>>> https://lists.wikimedia.org/mailman/listinfo/wikitech-l >>>>>> >>>>>>