On Mon, 10 Dec 2018 11:22:01 +0100, Takashi Iwai wrote:
On Sat, 08 Dec 2018 02:55:43 +0100, Andre Guedes wrote:
Hi all,
This patch series provides some follow-up improvements to the AVTP Audio Format (AAF) plugin. The highlight of this series is the implementation of a transmission offload mechanism which improves the plugin performance considerably. Details are provided in the following paragraphs.
Currently, the AVTPDU transmission interval is controlled by the AAF plugin, in software. At every timer expiration, the plugin transmits one AVTPDU. This has some implications in terms of task scheduling. For instance, in class A streams, AVTPDUs are transmitted at every 125 us which means the application task should be scheduled-in at every 125 us. In a general purpose Linux system, such scheduling interval can be hard to cope with in the long-run.
To mitigate that issue, this series introduces a transmission offload mechanism that leverages the SO_TXTIME sockopt and ETF qdisc features recently introduced to kernel 4.19. Instead of sending one AVTPDU at every timer expiration, the plugin sends several AVTPDUs at once to the kernel, configuring their Tx time so the transmission interval is maintained. The kernel can then offload packet transmission to the hardware (if the network controller supports it), providing more transmission time accuracy. This offloading mechanism enables the application to sleep for longer times, easing on task scheduling.
To illustrate the improvements provided by this series, I ran a before-after experiment. The experiment setup consisted in 2 PCs with Intel i210 card connected back-to-back running an up-to-date Archlinux with kernel 4.19.4. I ran 'aplay' for 5 minutes on one PC and captured the AVTPDUs on the other PC. The metric under evaluation is the transmission interval and it is measured by checking the 'time_delta' information from ethernet frames captured at the receiving side. If you're interested in reproducing the experiments, let me know and I can share my helper scripts.
The table below shows the experiment outcome for a Class A, stereo, 16-bit sample, 48 kHz stream. The unit is nanoseconds.
| Mean | Stdev | Min | Max | Range |-------+--------+---------+---------+---------+---------+ Before | 125000 | 1154 | 75311 | 172144 | 96833 | After | 125000 | 18 | 124960 | 125048 | 88 |
Before this patchset, the transmission interval mean is equal to the optimal value (Class A stream -> 125 us interval), and it is kept the same after the patchset. However, the dispersion measurements had improved considerably, meaning the system is consistently transmitting AVTPDUs at the correct interval.
Finally, to help the review process, here follows a quick summary of the patches within this series:
- PATCH 1: fixes a type in the plugin documentation.
- PATCH 2: enables the user to configure the presentation time tolerance.
- PATCH 3-5: refactor the code in order to land the offload mechanism code smoothly.
- PATCH 6-7: implement the transmission offload mechanism.
This series can also be found in my alsa-plugins tree in github [1].
Regards,
Andre
[1] https://github.com/aguedes/alsa-plugins
Andre Guedes (7): doc: Fix typo in AAF doc aaf: Add presentation time tolerance aaf: Refactor AVTPDU transmission routines aaf: Refactor AVTPDU reception routines aaf: Refactor timeout routines aaf: Tx multiple AVTPDUs per media clock tick aaf: AVTPDU transmission periodicity
Applied all patches now. Thanks.
BTW, while building libavtp for 32bit x86, the unit test failed: The build log is found at https://build.opensuse.org/package/live_build_log/home:tiwai/libavtp/openSUS...
x86_64 seems working, so it must be specific to 32bit arch.
Takashi