Subject: [linux-audio-dev] Fwd: [rtl] Beowulf and RTLinux (and NIC latency)
From: David Olofson (audiality_AT_swipnet.se)
Date: ti syys 28 1999 - 18:57:27 EDT
For those interested in real time processing using Beowulf clusters and similar
solutions; here's a post from the RTLinux list...
The point: I certainly don't see a significant hardware problem with real time
audio processing on clusters.
An engine that keeps track of inter-plug-in dependencies shouldn't have much
trouble making use of a Beowulf. There would be very few points in the net
where the network transfer latency come in, so the sub nets on the nodes would
be running nicely in parallel, with a time offset that depends on the transfer
latency/bandwidth ratio.
What's a few ms (if even that much) extra per node crossing (one or two in the
normal case) for mix down processing and playing from the sequencer? People are
actually *using* Windoze for that kind of work... (*muhahaha...!* Sorry. I
sure know what it feels like. ;-)
...so for the post:
---------------------------------------------------------
From: David Schleef <ds_AT_stm.lbl.gov>
To: David Olofson <david.olofson_AT_reologica.se>
Cc: yodaiken_AT_fsmlabs.com,
"Cesar A.R.G" <cero_AT_beta.sucre.udo.edu.ve>,
Beowulf <beowulf_AT_beowulf.gsfc.nasa.gov>,
RTLinux <rtl_AT_rtlinux.cs.nmt.edu>
Subject: Re: [rtl] Beowulf and RTLinux (and NIC latency)
---------------------------------------------------------
On Mon, Sep 27, 1999 at 04:23:45PM +0200, David Olofson wrote:
> I'd like to see some more figures on actual transfer rates and latencies
> that
> can be achieved with RTLinux drivers.
Here's some latency numbers for the RT networking stuff that I'm
doing. The numbers aren't very good, but the code is in the "Make
It Work Right" stage, and not tuned.
Computer 1: Pentium 90, 3c905b 100baseT NIC, Linux-2.2.10-rtl11
Computer 2: dual-P150, 3c905b 100baseT NIC, Linux-2.2.10-rtl11 (running
single processor, 'cuz it crashed otherwise)
connected by a 100baseT crossover cable
I'm getting round-trip times for ~60 byte UDP packets of consistently
less than 500 us for 100 million packets over 14 hours. This number is
from RT task to RT task on the other computer and back using a simple
UDP echo server (port 7).
I have good reason to believe that with faster computers, the round trip
time should be less than 100 us. Theoretically, it could be even as the
25-50 us range, but this might require significant hacks. I get ping times
between a PII-400 and PPro200 (i.e., faster computers) of 0.1 ms with
normal Linux networking.
As for bandwidth, haven't tested it.
dave...
---------------------------------------------------------
//David
·A·U·D·I·A·L·I·T·Y· P r o f e s s i o n a l L i n u x A u d i o
- - ------------------------------------------------------------- - -
·Rock Solid David Olofson:
·Low Latency www.angelfire.com/or/audiality ·Audio Hacker
·Plug-Ins audiality_AT_swipnet.se ·Linux Advocate
·Open Source ·Singer/Composer
This archive was generated by hypermail 2b28 : pe maalis 10 2000 - 07:27:12 EST