On Sat, 14 Mar 2020 09:11:17 +0100, Andreas Steinmetz wrote:
The following patch series fixes USB MIDI output starvation and allows to configure a predictable worst case output latency with a minimum of 15ms for 16x16 devices and 3.8ms for 4x4 devices (excluding USB transfer time and latencies of the interface device itself).
The reasoning behind this patch series is first to fix the possible indefinite latency and then to allow to configure a predictable low latency to be able to use a cheap embedded Linux system (e.g. Raspberry Pi) in combination with a multiport USB MIDI interface as a drop in replacement for commercial hardware based MIDI routing and filtering products. This includes "real time" (live) usage.
For this usage the guaranteed latency within the kernel must be as low as possible, depending on the requirements of the USB MIDI interface in use.
Though the current state of the kernel code is acceptable in a fully controlled by a single user environment, it is not usable for uncontrolled standalone use as the driver behaviour then depends on unpredictable MIDI data streams and thus is unpredictable in behaviour itself.
After applying this patch series the USB MIDI driver can be used as usual or it can be tuned to low latency when this is desired.
From a developer perspective the actual output rate of USB MIDI
interfaces is typically slower than the expected 320us and even varies depending on how many ports are busy (e.g. a M-Audio Midisport 4x4 AE interface exhibits a varying output rate of 357us to 515us per byte depending on device load). Thus the driver needs a predictable low latency to allow userspace to use the rawmidi buffer fill state to estimate the expected output latency and schedule data accordingly.
The following table shows relative processing time for varying driver configurations using the M-Audio Midisport 4x4 AE as the test device. For the parameters "outqueue" and "events" do see the descriptions of the following patches.
| unpatched | outqueue=7 | outqueue=1 | outqueue=1 | | kernel | events=0 | events=0 | events=4 | --------+---------------+---------------+---------------+---------------+ 1 Port | 100% | 100% | 100% | 100.05% | --------+---------------+---------------+---------------+---------------+ 4 Ports | n/a | 100% | 99.97% | 100.36% | --------+---------------+---------------+---------------+---------------+
As can be seen from the above table the changes to output latency do affect throughput only marginally.
Thanks for the patches and the analysis.
So, judging from the experiments above, outqueue=1 has no measurable performance regression? For 4 ports, it's even an improvement. Then we may change the default value.
I guess you kept the old value just to be conservative, but if the number tells a better story and it's no part or ABI, we may consider changing the default for a better performance.
Or, is there any other potential risk by changing this?
thanks,
Takashi