On 04/02/2016 06:33 PM, Jonathan E. Brickman wrote:
For those who may be interested, I am seeing the first
small glimmering
of what looks like real success, in running multiple JACK servers on one
box for the purpose of distributing DSP load. I can now reliably have
two netjack1 slaves connect to one netjack1 master, the master running
ALSA to hardware, all within one Linux box under one kernel and one
filesystem.
The way that this works is "containerization" or "sandboxing", which
can
give an IP to anything ranging from one running binary to a large
collection. I started the current effort using Docker containerization,
but its standard setups require a lot of filesystem recreation, each
Docker container has an entire working OS filesystem apart from the
kernel; but during the studies for Docker, I blundered into something
else called firejail, and firejail appears to be just what the doctor
ordered, it seems to do just about everything while using the existing
filesystem, and one can very easily turn off the features one doesn't need.
To cut to the chase, after boot and system prep, first one creates an IP
bridge:
sudo brctl addbr br0
sudo ifconfig br0 10.99.99.1/24
and then starts JACK in the sandboxes:
# The master hardware sandbox is named MASTER, the QJackCTL profile
is named MASTER,
# the JACK server is named MASTER.
nohup firejail --name=MASTER --noprofile --net=br0 --ip=10.99.99.10 \
/usr/bin/jackd -n MASTER -m -dalsa -r48000 -p256 -n2 -Xseq -D
-Chw:NVidia -Phw:NVidia \
~/LOGS/jack-master.log &
# The SRO hardware sandbox is named SRO, the QJackCTL profile for it
is named SRO,
# the JACK server is named SRO.
nohup firejail --name=SRO --noprofile --net=br0 --ip=10.99.99.15 \
/usr/bin/jackd -n SRO -dnetone \
~/LOGS/jack-sro.log &
# Ditto for STRINGS.
nohup firejail --name=STRINGS --noprofile --net=br0 --ip=10.99.99.20 \
/usr/bin/jackd -n STRINGS -dnetone \
~/LOGS/jack-strings.log &
And then connects the JACK servers, by adding jack_netsource processes
into the existing MASTER sandbox:
#!/bin/bash
# SRO
nohup firejail --join=MASTER jack_netsource -s MASTER -H 10.99.99.15
~/LOGS/netsource-sro.log &
# STRINGS
nohup firejail --join=MASTER jack_netsource -s MASTER -H 10.99.99.20
~/LOGS/netsource-strings.log &
I have not tested further yet, no clients -- am working on something
closer to production now -- but this is looking very good, the
connections are reported successful, zero xruns, and on this prototyping
box -- 2.6GHz AMD quad, ten-plus years old -- 0.8% DSP in use on all
three JACK servers and very low memory usage. This is considerably
better than I had hoped.
The sandboxed binaries above, cannot reach outside the one box as
written above, but the excellent and very responsive developer of
Firejail has provided a method
<https://github.com/netblue30/firejail/issues/372>. As a result this
could all be done box-independently -- for instance, if one's "monolith
is the building" (going to have to remember that, Patrick), one could
have all indicated motherboards NFS or SMB to one file server, and then
use a central control GUI machine to run all firejails on whichever
hardware was proven most appropriate, and testing could become much
easier. Along the way I also found 'xpra' likely to be a very good way
to setup such a central control, am going to test that as a side project.
I tried using netjack2 first, but ran into mystery behavior, xruns
started piling huge when the second slave connected. So I went with
netjack1, especially because Patrick Shirkey already proved the above
paradigm in multibox mode using netjack1. I am currently using netjack1
under jackd2, but will change to netjack under jackd1 if a reason to do
so appears.
I am now working on a working dual-JACK prototype, in a near-production
design, as a next step towards a generally transportable MultiJACK Patch
Management methodology and the next big build of my Box of No Return :-)
Cheers, and thanks everyone!!!!
Hmmm... I wonder: this is either a fundamental bug in JACK2 that none of
the developers seem to be able to reproduce, a complete misunderstanding
on the part of the original poster, or something completely different
that is not sufficiently understood. In any case, I'd say it would be
useful to come up with at least a working hypothesis of why this should
improve things.
I have never used a graph as complex as Jonathan's, but I don't see how
adding virtual machine and networking overhead could improve throughput
when JACK2 should already parallelize everything that can be
parallelized... Wouldn't the time be better spent to analyze JACK2's
scheduling behaviour and catch bugs if there are any?
When exporting, I have seen something that might be similar to what
Jonathan might be seeing: a freewheeling, 100% CPU-bound process only
seems to use a fraction of the available CPU even on a single core. Now
this might indicate a real problem, or it might indicate that the CPU
load tools I'm using do not accurately report what's going on under
tight real-time conditions with loads of context switches.
--
Jörn Nettingsmeier
Lortzingstr. 11, 45128 Essen, Tel. +49 177 7937487
Meister für Veranstaltungstechnik (Bühne/Studio)
Tonmeister VDT