Posted on Di 08 April 2008

What's Cooking in PulseAudio's glitch-free Branch

A while ago I started development of special branch of PulseAudio which is called glitch-free. In a few days I will merge it back to PulseAudio trunk, and eventually release it as 0.9.11. I think it's time to explain a little what all this "glitch-freeness" is about, what made it so tricky to implement, and why this is totally awesome technology. So, here we go:

Traditional Playback Model

Traditionally on most operating systems audio is scheduled via sound card interrupts (IRQs). When an application opens a sound card for playback it configures it for a fixed size playback buffer. Then it fills this buffer with digital PCM sample data. And after that it tells the hardware to start playback. Then, the hardware reads the samples from the buffer, one at a time, and passes it on to the DAC so that eventually it reaches the speakers.

After a certain number of samples played the sound hardware generates an interrupt. This interrupt is forwarded to the application. On Linux/Unix this is done via poll()/select(), which the application uses to sleep on the sound card file descriptor. When the application is notified via this interrupt it overwrites the samples that were just played by the hardware with new data and goes to sleep again. When the next interrupt arrives the next block of samples is overwritten, and so on and so on. When the hardware reaches the end of the hardware buffer it starts from its beginning again, in a true ring buffer fashion. This goes on and on and on.

The number of samples after which an interrupt is generated is usually called a fragment (ALSA likes to call the same thing a period for some reason). The number of fragments the entire playback buffer is split into is usually integral and usually a power of two, 2 and 4 being the most frequently used values.

Schematic overview
Image 1: Schematic overview of the playback buffer in the traditional playback model, in the best way the author can visualize this with his limited drawing abilities.

If the application is not quick enough to fill up the hardware buffer again after an interrupt we get a buffer underrun ("drop-out"). An underrun is clearly hearable by the user as a discontinuity in audio which is something we clearly don't want. We thus have to carefully make sure that the buffer and fragment sizes are chosen in a way that the software has enough time to calculate the data that needs to be played, and the OS has enough time to forward the interrupt from the hardware to the userspace software and the write request back to the hardware.

Depending on the requirements of the application the size of the playback buffer is chosen. It can be as small as 4ms for low-latency applications (such as music synthesizers), or as long as 2s for applications where latency doesn't matter (such as music players). The hardware buffer size directly translates to the latency that the playback adds to the system. The smaller the fragment sizes the application configures, the more time the application has to fill up the playback buffer again.

Let's formalize this a bit: Let BUF_SIZE be the size of the hardware playback buffer in samples, FRAG_SIZE the size of one fragment in samples, and NFRAGS the number of fragments the buffer is split into (equivalent to BUF_SIZE divided by FRAG_SIZE), RATE the sampling rate in samples per second. Then, the overall latency is identical to BUF_SIZE/RATE. An interrupt is generated every FRAG_SIZE/RATE. Every time one of those interrupts is generated the application should fill up one fragment again, if it missed one interrupt this might become more than one. If it doesn't miss any interrupt it has (NFRAGS-1)*FRAG_SIZE/RATE time to fulfill the request. If it needs more time than this we'll get an underrun. The fill level of the playback buffer should thus usually oscillate between BUF_SIZE and (NFRAGS-1)*FRAG_SIZE. In case of missed interrupts it might however fall considerably lower, in the worst case to 0 which is, again, an underrun.

It is difficult to choose the buffer and fragment sizes in an optimal way for an application:

  • The buffer size should be as large as possible to minimize the risk of drop-outs.
  • The buffer size should be as small as possible to guarantee minimal latencies.
  • The fragment size should be as large as possible to minimize the number of interrupts, and thus the required CPU time used, to maximize the time the CPU can sleep for between interrupts and thus the battery lifetime (i.e. the fewer interrupts are generated the lower your audio app will show up in powertop, and that's what all is about, right?)
  • The fragment size should be as small as possible to give the application as much time as possible to fill up the playback buffer, to minimize drop-outs.

As you can easily see it is impossible to choose buffering metrics in a way that they are optimal on all four requirements.

This traditional model has major drawbacks:

  • The buffering metrics are highly dependant on what the sound hardware can provide. Portable software needs to be able to deal with hardware that can only provide a very limited set of buffer and fragment sizes.
  • The buffer metrics are configured only once, when the device is opened, they usually cannot be reconfigured during playback without major discontinuities in audio. This is problematic if more than one application wants to output audio at the same time via a sound server (or dmix) and they have different requirements on latency. For these sound servers/dmix the fragment metrics are configured statically in a configuration file, and are the same during the whole lifetime. If a client connects that needs lower latencies, it basically lost. If a client connects that doesn't need as low latencies, we will continouisly burn more CPU/battery than necessary.
  • It is practically impossible to choose the buffer metrics optimal for your application -- there are too many variables in the equation: you can't know anything about the IRQ/scheduling latencies of the OS/machine your software will be running on; you cannot know how much time it will actually take to produce the audio data that shall be pushed to the audio device (unless you start counting cycles, which is a good way to make your code unportable); the scheduling latencies are hugely dependant on the system load on most current OSes (unless you have an RT system, which we generally do not have). As said, for sound servers/dmix it is impossible to know in advance what the requirements on latency are that the applications that might eventually connect will have.
  • Since the number of fragments is integral and at least 2 on almost all existing hardware we will generate at least two interrupts on each buffer iteration. If we fix the buffer size to 2s then we will generate an interrupt at least every 1s. We'd then have 1s to fill up the buffer again -- on all modern systems this is far more than we'd ever need. It would be much better if we could fix the fragment size to 1.9s, which still gives us 100ms to fill up the playback buffer again, still more than necessary on most systems.

Due to the limitations of this model most current (Linux/Unix) software uses buffer metrics that turned out to "work most of the time", very often they are chosen without much thinking, by copying other people's code, or totally at random.

PulseAudio <= 0.9.10 uses a fragment size of 25ms by default, with four fragments. That means that right now, unless you reconfigure your PulseAudio manually clients will not get latencies lower than 100ms whatever you try, and as long as music is playing you will get 40 interrupts/s. (The relevant configuration options for PulseAudio are default-fragments= and default-fragment-size-msec= in daemon.conf)

dmix uses 16 fragments by default with a size of 21 ms each (on my system at least -- this varies, depending on your hardware). You can't get less than 47 interrupts/s. (You can change the parameters in .asoundrc)

So much about the traditional model and its limitations. Now, we'll have a peek on how the new glitch-free branch of PulseAudio does its things. The technology is not really new. It's inspired by what Vista does these days and what Apple CoreAudio has already been doing for quite a while. However, on Linux this technology is new, we have been lagging behind quite a bit. Also I claim that what PA does now goes beyond what Vista/MacOS does in many ways, though of course, they provide much more than we provide in many other ways. The name glitch-free is inspired by the term Microsoft uses to call this model, however I must admit that I am not sure that my definition of this term and theirs actually is the same.

Glitch-Free Playback Model

The first basic idea of the glitch-free playback model (a better, less marketingy name is probably timer-based audio scheduling which is the term I internally use in the PA codebase) is to no longer depend on sound card interrupts to schedule audio but use system timers instead. System timers are far more flexible then the fragment-based sound card timers. They can be reconfigured at any time, and have a granularity that is independant from any buffer metrics of the sound card. The second basic idea is to use playback buffers that are as large as possible, up to a limit of 2s or 5s. The third basic idea is to allow rewriting of the hardware buffer at any time. This allows instant reaction on user-input (i.e. pause/seek requests in your music player, or instant event sounds) although the huge latency imposed by the hardware playback buffer would suggest otherwise.

PA configures the audio hardware to the largest playback buffer size possible, up to 2s. The sound card interrupts are disabled as far as possible (most of the time this means to simply lower NFRAGS to the minimal value supported by the hardware. It would be great if ALSA would allow us to disable sound card interrupts entirely). Then, PA constantly determines what the minimal latency requirement of all connected clients is. If no client specified any requirements we fill up the whole buffer all the time, i.e. have an actual latency of 2s. However, if some applications specified requirements, we take the lowest one and only use as much of the configured hardware buffer as this value allows us. In practice, this means we only partially fill the buffer each time we wake up. Then, we configure a system timer to wake us up 10ms before the buffer would run empty and fill it up again then. If the overall latency is configured to less than 10ms we wakeup after half the latency requested.

If the sleep time turns out to be too long (i.e. it took more than 10ms to fill up the hardware buffer) we will get an underrun. If this happens we can double the time we wake up before the buffer would run empty, to 20ms, and so on. If we notice that we only used much less than the time we estimated, we can halve this value again. This adaptive scheme makes sure that in the unlikely event of a buffer underrun it will happen most likely only once and never again.

When a new client connects or an existing client disconnects, or when a client wants to rewrite what it already wrote, or the user wants to change the volume of one of the streams, then PA will resample its data passed by the client, convert it to the proper hardware sample type, and remix it with the data of the other clients. This of course makes it necessary to keep a "history" of data of all clients around so that if one client requests a rewrite we have the necessary data around to remix what already was mixed before.

The benefits of this model are manyfold:

  • We minimize the overall number of interrupts, down to what the latency requirements of the connected clients allow us. i.e. we save power, don't show up in powertop anymore for normal music playback.
  • We maximize drop-out safety, because we buffer up to 2s in the usual cases. Only with operating systems which have scheduling latencies > 2s we can still get drop-outs. Thankfully no operating system is that bad.
  • In the event of an underrun we don't get stuck in it, but instead are able to recover quickly and can make sure it doesn't happen again.
  • We provide "zero-latency". Each client can rewrite its playback buffer at any time, and this is forwarded to the hardware, even if this means that the sample currently being played needs to be rewritten. This means much quicker reaction to user input, a more responsive user experience.
  • We become much less dependant on what the sound hardware provides us with. We can configure wakeup times that are independant from the fragment settings that the hardware actually supports.
  • We can provide almost any latency a client might request, dynamically without reconfiguration, without discontinuities in audio.

Of course, this scheme also comes with major complications:

  • System timers and sound card timers deviate. On many sound cards by quite a bit. Also, not all sound cards allow the user to query the playback frame index at any time, but only shortly after each IRQ. To compensate for this deviation PA contains a non-trivial algorithm which tries to estimate and follow the deviation over time. If this doesn't work properly it might happen that an underrun happens much earlier than we expected.
  • System timers on Unix are not very high precision. On traditional Linux with HZ=100 sleep times for timers are rounded up to multiples of 10ms. Only very recent Linux kernels with hrtimers can provide something better, but only on x86 and x86-64 until now. This makes the whole scheme unusable for low latency setups unless you run the very latest Linux. Also, hrtimers are not (yet) exposed in poll()/select(). It requires major jumping through loops to work around this limitation.
  • We need to keep a history of sample data for each stream around, thus increasing the memory footprint and potentially increased cache pressure. PA tries to work against the increased memory footprint and cache pressure this might cause by doing zero-copy memory management.
  • We're still dependant on the maximum playback buffer size the sound hardware supports. Many sound cards don't even support 2s, but only 300ms or suchlike.
  • The rewriting of the client buffers causing rewriting of the hardware buffer complicates the resampling/converting step immensly. In general the code to implement this model is more complex than for the traditional model. Also, ALSA has not really been designed with this design in mind, which makes some things very hard to get right and suboptimal.
  • Generally, this works reliably only on newest ALSA, newest kernel, newest everything. It has pretty steep requirements on software and sometimes even on hardware. To stay comptible with systems that don't fulfill these requirements we need to carry around code for the traditional playback model as well, increasing the code base by far.

The advantages of the scheme clearly outweigh the complexities it causes. Especially the power-saving features of glitch-free PA should be enough reason for the embedded Linux people to adopt it quickly. Make PA disappear from powertop even if you play music!

The code in the glitch-free is still rough and sometimes incomplete. I will merge it shortly into trunk and then upload a snapshot to Rawhide.

I hope this text also explains to the few remaining PA haters a little better why PA is a good thing, and why everyone should have it on his Linux desktop. Of course these changes are not visible on the surface, my hope with this blog story is to explain a bit better why infrastructure matters, and counter misconceptions what PA actually is and what it gives you on top of ALSA.

© Lennart Poettering. Built using Pelican. Theme by Giulio Fidente on github. .