On Tue, 30 Mar 2021 21:35:11 +0200, David Henningsson wrote:
On 2021-03-28 09:40, Takashi Iwai wrote:
On Sun, 28 Mar 2021 08:39:46 +0200, David Henningsson wrote:
Hi Takashi and Takashi,
You both question the usability of the patch, so let's take a step back.
Suppose you're writing the next JACK, or a DAW, or something like that. When writing a DAW, you need to support the users who need ultra-low latency for live playing of an instrument. These users (unfortunately) need to reconfigure their Linux installation, have special kernels, buy expensive sound cards etc, in order to get the best possible latency. You also should give the best possible experience for people who don't have the time to do that. Just recording a simple MIDI file should not require any extra kernel options, RT_PRIO privileges or anything like that. (And then there are people in between, who try to get the best possible latency given their limited time, money and skills.)
Now you're asking yourself whether to use rawmidi or seq API. It seems silly to have to support both. The seq interface is suboptimal for the first use case, due to the latency introduced by the workqueue. But rawmidi is entirely impossible for the second use case, due to the lack of timestamping. (From a quick look at Ardour's sources, it does support both rawmidi and seq. The rawmidi code mostly timestamps the message and sends it to another thread. [1] I e, essentially what I believe the kernel should do, because that timestamp is better.)
What you don't need is exact measurements of burst interval or even timestamp accuracy. All you have use for is the best possible timestamp, because that's what's going to be written into the MIDI file. There might be other use cases for burst intervals etc, but I don't see them helping here.
On 2021-03-26 17:44, Takashi Iwai wrote:
On Fri, 26 Mar 2021 17:29:04 +0100, David Henningsson wrote:
But actually I'd like to see some measurement how much we can improve the timestamp accuracy by shifting the post office. This may show interesting numbers.
Sorry, I don't know the idiom "shifting the post office" and neither does the urban dictionary, so I have no idea what this means. :-)
It was just joking; you basically moved the place to stamp the incoming data from one place (at the delivery center of a sequencer event) to another earlier place (at the irq handler).
The question is: how much time difference have you measured by this move?
Ok, thanks for the explanation. I have not done any measurements because it would be quite time consuming to do so, across different hardware, kernel configurations, and so on. I don't have that time right now, sorry. But the claim that workqueues can be delayed up to a second (!) from just adding a few RT_PRIO tasks [2] is enough to scare me from using the seq interface for accurate timestamping.
Also, one thing to be more considered is the support for MIDI v2 in future. I haven't seen any development so far (and no device available around), so I cannot comment on this much more, but it'd be worth to take a quick look before defining the solid API/ABI.
I had a quick look at MIDI 2.0. It offers something called "Jitter reduction timestamps". After some searching I found that its resolution is 16 bit, and in units of 1/31250 seconds [1]. So the suggested timestamp format of secs + nsecs would suit us well for that case, I believe. When implemented, MIDI 2.0 jitter reduction timestamps would be another clock ID on top of the existing frame format (or a new frame format, if we prefer).
A midi 2.0 UMP (Universal Midi Packet) seems to be 4, 8, 12 or 16 bytes, excluding the timestamp. If we want to fit that format with the existing patch, we could increase the frame to 32 bytes so we can fit more data per packet. Do you think we should do that? Otherwise I think Patch v3 is ready for merging.
Let's evaluate a bit what would be the best fit. I see no big reason to rush the merge right now.
Does this mean "evaluate for a week or two because of kernel cadence, merge windows etc" or does this mean "evaluate for months or years until someone does a full MIDI 2.0 kernel implementation"?
Well, without the actual measurement, it's purely a theoretical problem, and it implies that we haven't seen any real improvement by that, too. So, the first priority is to measure and prove the need of the changes.
Well, I believe that rawmidi provides less jitter than seq is not a theoretical problem but a known fact (see e g [1]), so I haven't tried to "prove" it myself. And I cannot read your mind well enough to know what you would consider a sufficient proof - are you expecting to see differences on a default or RT kernel, on a Threadripper or a Beaglebone, idle system or while running RT torture tests? Etc.
There is certainly difference, and it might be interesting to see the dependency on the hardware or on the configuration. But, again, my primary question is: have you measured how *your patch* really provides the improvement? If yes, please show the numbers in the patch description.
That said; there are certainly people who run the seq interface succesfully as well. It depends on both hardware, drivers, system load, kernel configuration etc (and perhaps also the timing skill of the musician!) if that work in the workqueue will be delayed often enough to not go unnoticed.
Let me ask a counter question. Suppose you were to write the next JACK, DAW as I wrote about above, where you need both the best possible latency and best possible timestamps. Would you as a maintainer recommend seq or rawmidi (the latter with timestamps taken from userspace)?
Sorry, you missed the whole point. The patch was written for the latency improvement, and it's a kind of performance stuff. For such a performance change, the major interest is the number the patch provides (improves). That's why I asked it. Without numbers, it's nothing but advertising myths.
I don't have any strong opinion about sequencer vs rawmidi usage. If any of them brings a better performance, that's fine and why not. We may improve the implementation of sequencer stuff if it's requested, too; but all needs to be done per measurement at first (if 4-5ms latency is found, it's really awful and worth for investigation).
Then the next thing is to determine the exact format for the new API in a solid form. It's still not fully agreed which frame size fits at best, for example. Also, we may have two individual frame types, e.g. a timestamp frame and a data frame, too, depending on the frame size and the implementation. And, it might be handy if the ioctl returns the frame size to user-space, too.
And, of course, thinking on MIDI 2.0 wouldn't be bad. Though I don't think tying with MIDI 2.0 is needed right now; instead, we should assure only that the new timestamp would be accurate enough for new extensions like MIDI 2.0.
Okay; I think we should then go for a frame size of 32 bytes with 16 byte header/timestamp and 16 byte data. One type of frame only, no frame size ioctl will be needed because any changes to the frame format would need a new framing type. I envision the application reading this struct directly without "encapsulation" with accessors in alsa-lib. This is for MIDI 2.0 compatibility: from what I can read, MIDI 2.0 messages can be up to 16 bytes. Its "Jitter Reduction Timestamps" are 2 bytes.
This sounds promising as future-proof, yeah.
thanks,
Takashi