For one of our customers at Centricular we were working on a quite interesting project. Their use-case was basically to receive an as-high-as-possible number of audio RTP streams over UDP, transcode them, and then send them out via UDP again. Due to how GStreamer usually works, they were running into some performance issues.
This blog post will describe the first set of improvements that were implemented for this use-case, together with a minimal benchmark and the results. My colleague Mathieu will follow up with one or two other blog posts with the other improvements and a more full-featured benchmark.
The short version is that CPU usage decreased by about 65-75%, i.e. allowing 3-4x more streams with the same CPU usage. Also parallelization works better and usage of different CPU cores is more controllable, allowing for better scalability. And a fixed, but configurable number of threads is used, which is independent of the number of streams.
The code for this blog post can be found here.
Table of Contents
- GStreamer & Threads
- Thread-Sharing GStreamer Elements
- Available Elements
- Little Benchmark
GStreamer & Threads
In GStreamer, by default each source is running from its own OS thread. Additionally, for receiving/sending RTP, there will be another thread in the RTP jitterbuffer, yet another thread for receiving RTCP (another source) and a last thread for sending RTCP at the right times. And RTCP has to be received and sent for the receiver and sender side part of the pipeline, so the number of threads doubles. In the sum this gives at least 1 + 1 + (1 + 1) * 2 = 6 threads per RTP stream in this scenario. In a normal audio scenario, there will be one packet received/sent e.g. every 20ms on each stream, and every now and then an RTCP packet. So most of the time all these threads are only waiting.
Apart from the obvious waste of OS resources (1000 streams would be 6000 threads), this also brings down performance as all the time threads are being woken up. This means that context switches have to happen basically all the time.
To solve this we implemented a mechanism to share threads, and in the end as a result we have a fixed, but configurable number of threads that is independent from the number of streams. And can run e.g. 500 streams just fine on a single thread with a single core, which was completely impossible before. In addition we also did some work to reduce the number of allocations for each packet, so that after startup no additional allocations happen per packet anymore for buffers. See Mathieu’s upcoming blog post for details.
In this blog post, I’m going to write about a generic mechanism for sources, queues and similar elements to share their threads between each other. For the RTP related bits (RTP jitterbuffer and RTCP timer) this was not used due to reuse of existing C codebases.
Thread-Sharing GStreamer Elements
The code in question can be found here, a small benchmark is in the examples directory and it is going to be used for the results later. A full-featured benchmark will come in Mathieu’s blog post.
This is a new GStreamer plugin, written in Rust and around the Tokio crate for asynchronous IO and generally a “task scheduler”.
While this could certainly also have been written in C around something like libuv, doing this kind of work in Rust is simply more productive and fun due to its safety guarantees and the strong type system, which definitely reduced the amount of debugging a lot. And in addition “modern” language features like closures, which make working with futures much more ergonomic.
When using these elements it is important to have full control over the pipeline and its elements, and the dataflow inside the pipeline has to be carefully considered to properly configure how to share threads. For example the following two restrictions should be kept in mind all the time:
- Downstream of such an element, the streaming thread must never ever block for considerable amounts of time. Otherwise all other elements inside the same thread-group would be blocked too, even if they could do any work now
- This generally all works better in live pipelines, where media is produced in real-time and not as fast as possible
So this repository currently contains the generic infrastructure (see the src/iocontext.rs source file) and a couple of elements:
- an UDP source: ts-udpsrc, a replacement for udpsrc
- an app source: ts-appsrc, a replacement for appsrc to inject packets into the pipeline from the application
- a queue: ts-queue, a replacement for queue that is useful for adding buffering to a pipeline part. The upstream side of the queue will block if not called from another thread-sharing element, but if called from another thread-sharing element it will pause the current task asynchronously. That is, stop the upstream task from producing more data.
- a proxysink/src element: ts-proxysrc, ts-proxysink, replacements for proxysink/proxysrc for connecting two pipelines with each other. This basically works like the queue, but split into two elements.
- a tone generator source around spandsp: ts-tonesrc, a replacement for tonegeneratesrc. This also contains some minimal FFI bindings for that part of the spandsp C library.
All these elements have more or less the same API as their non-thread-sharing counterparts.
API-wise, each of these elements has a set of properties for controlling how it is sharing threads with other elements, and with which elements:
- context: A string that defines in which group this element is. All elements with the same context are running on the same thread or group of threads,
- context-threads: Number of threads to use in this context. -1 means exactly one thread, 1 and above used N+1 threads (1 thread for polling fds, N worker threads) and 0 sets N to the number of available CPU cores. As long as no considerable work is done in these threads, -1 has shown to be the most efficient. See also this tokio GitHub issue
- context-wait: Number of milliseconds that the threads will wait on each iteration. This allows to reduce CPU usage even further by handling all events/packets that arrived during that timespan to be handled all at once instead of waking up the thread every time a little event happens, thus reducing context switches again
The elements are all pushing data downstream from a tokio thread whenever data is available, assuming that downstream does not block. If downstream is another thread-sharing element and it would have to block (e.g. a full queue), it instead returns a new future to upstream so that upstream can asynchronously wait on that future before producing more output. By this, back-pressure is implemented between different GStreamer elements without ever blocking any of the tokio threads. All this is implemented around the normal GStreamer data-flow mechanisms, there is no “tokio fast-path” between elements.
As mentioned above, there’s a small benchmark application in the examples directory. This basically sets up a configurable number of streams and directly connects them to a fakesink, throwing away all packets. Additionally there is another thread that is sending all these packets. As such, this is really the most basic benchmark and not very realistic but nonetheless it shows the same performance improvement as the real application. Again, see Mathieu’s upcoming blog post for a more realistic and complete benchmark.
When running it, make sure that your user can create enough fds. The benchmark will just abort if not enough fds can be allocated. You can control this with ulimit -n SOME_NUMBER, and allowing a couple of thousands is generally a good idea. The benchmarks below were running with 10000.
After running cargo build –release to build the plugin itself, you can run the benchmark with:
cargo run --release --example udpsrc-benchmark -- 1000 ts-udpsrc -1 1 20
and in another shell the UDP sender with
cargo run --release --example udpsrc-benchmark-sender -- 1000
This runs 1000 streams, uses ts-udpsrc (alternative would be udpsrc), configures exactly one thread -1, 1 context, and a wait time of 20ms. See above for what these settings mean. You can check CPU usage with e.g. top. Testing was done on an Intel i7-4790K, with Rust 1.25 and GStreamer 1.14. One packet is sent every 20ms for each stream.
Results for 3000 streams for the old udpsrc are not included as starting up that many threads needs too long.
The best configuration is apparently a single thread per context (see this tokio GitHub issue) and waiting 20ms for every iterations. Compared to the old udpsrc, CPU usage is about one third in that setting, and generally it seems to parallelize well. It’s not clear to me why the last test has 11% more CPU with two contexts, while in every other test the number of contexts does not really make a difference, and also not for that many streams in the real test-case.
The waiting does not reduce CPU usage a lot in this benchmark, but on the real test-case it does. The reason is most likely that this benchmark basically sends all packets at once, then waits for the remaining time, then sends the next packets.
Take these numbers with caution, the real test-case in Mathieu’s blog post will show the improvements in the bigger picture, where it was generally a quarter of CPU usage and almost perfect parallelization when increasing the number of contexts.
Generally this was a fun exercise and we’re quite happy with the results, especially the real results. It took me some time to understand how tokio works internally so that I can implement all kinds of customizations on top of it, but for normal usage of tokio that should not be required and the overall design makes a lot of sense to me, as well as the way how futures are implemented in Rust. It requires some learning and understanding how exactly the API can be used and behaves, but once that point is reached it seems like a very productive and performant solution for asynchronous IO. And modelling asynchronous IO problems based on the Rust-style futures seems a nice and intuitive fit.
The performance measurements also showed that GStreamer’s default usage of threads is not always optimal, and a model like in upipe or pipewire (or rather SPA) can provide better performance. But as this also shows, it is possible to implement something like this on top of GStreamer and for the common case, using threads like in GStreamer reduces the cognitive load on the developer a lot.
For a future version of GStreamer, I don’t think we should make the threading “manual” like in these two other projects, but instead provide some API additions that make it nicer to implement thread-sharing elements and to add ways in the GStreamer core to make streaming threads non-blocking. All this can be implemented already, but it could be nicer.
All this “only” improved the number of threads, and thus the threading and context switching overhead. Many other optimizations in other areas are still possible on top of this, for example optimizing receive performance and reducing the number of memory copies inside the pipeline even further. If that’s something you would be interested in, feel free to get in touch.
And with that: Read Mathieu’s upcoming blog posts about the other parts, RTP jitterbuffer / RTCP timer thread sharing, and no allocations, and the full benchmark.
One thought on “Improving GStreamer performance on a high number of network streams by sharing threads between elements with Rust’s tokio crate”