Re: [Jack-Devel] Very low latency audio

Maybe I'm mis-understanding the device, but you can generally do < 64 byte DMA transfers over PCIe ... they are called partials. They aren't particularly efficient from a bus and memory utilization perspective ..  you start to see the inefficiency as you approach back-to-back transfers -> 1 usec (at least on current CPUs ... some 6+ yr old laptop its probably in the 10-20 usec area ... close to a 48KHz sample rate ... ).

The current state of the art for low-latency *anything* is to avoid interrupts (as an interrupt itself is ~1-2 usec of processing latency) - either in-kernel (e.g. 3.11's low-latency sockets) or in user space.

Eric Mann / Intel Networking Division

From: [hidden] [mailto:[hidden]] On Behalf Of Paul Davis
Sent: Friday, September 20, 2013 6:58 AM
To: John Reed
Cc: JACK
Subject: Re: [Jack-Devel] Very low latency audio

note that the lower limit for PCI burst transfer is 64 bytes, which means that if you are only schlepping around two channels worth of 16 bits each, you need 16 samples minimum to be able to launch a DMA transfer.

obviously higher channel counts reduce or eliminate this problem.

On Fri, Sep 20, 2013 at 8:35 AM, John Reed <[hidden]<mailto:[hidden]>> wrote:
Hello,

On my netbook, I was able to start jackd with
jackd -P62 -dalsa -dhw:0 -r44100 -p16 -n2
but the hardware driver didn't support -p8 or -p4.

-p16 already requires a lot of fast interrupts, so I assume that to use this reliably, I would have to be very careful about setting all of the kernel switches, etc.

Is there any hope for getting -p8 or -p4 working, even if I have to write my own custom device driver? Or has anyone done any work in connecting xenomai to jack? Would I be better off writing a xenomai application and then find a way to connect it to an ALSA driver directly?

Cheers,
John


[hidden]<mailto:[hidden]>

1379688134.19455_0.ltw:2, <4E396B9AA334024C8EF7B8EDD7B32D4557DF2FE9 at ORSMSX101 dot amr dot corp dot intel dot com>