Cesar A.R.G wrote:
>
> It is possible to implement a system of control in real time using Beowulf?
Some months ago, i had a discussion with somebody at the Cluster-Computing
Center
at the university of Paderborn/Germany on that topic.
They have clustered their 96 computer with SCI (scalable coherent interface -
that provides some sort of distributed shared memory - the setup of the
beast will be done by the standard linux-SCI-drivers, so rt-linux doesn't
need to care about that).
So RT-Linux doesn't need to know about clustering at all, i.e. 96 RT-Tasks
are running on the 96 equal equipt machines.
So there could be a rather simple algorithm:
Automatic Syncronisation
CPU0 ...---|-- 1.) Calculation ---|--- 2.) Data tranfer --|-- 3.) Sleep
------|-- 1) ...
CPU1 ...---|-- 1.) Calculation -|--- 2.) Data tranfer ------|-- 3.) Sleep
----|-- 1) ...
CPU2 ...---|-- 1.) Calculation ----|--- 2.) Data tranfer --|-- 3.) Sleep
-----|-- 1) ...
CPU3 ...---|-- 1.) Calculation ---|--- 2.) Data tranfer --|-- 3.) Sleep
------|-- 1) ...
|<-------- 40-70% ------|----- ~20% ------------|----- 10-40%
-----|-- 1)
|<--------- one control-cycle, maybe i.e. 1 or 10 milliseconds
------|--
^ ^
| <--------- all Tasks have the same wakeup-time ----------------- |
1.) Calcualtion (data sampling etc.) may take 40-70% of the CPU-time
2.) Data Transfer depends on SCI-latencys, but shouldn't take more then 20% of a
complete cycle
3.) Sleeping RT-Tasks mean that Linux has some time to do something (10-40%)
Even when the 96 CPUs are only used in 40% of the CPU-time for calculation,
that is a rather good value for many applications, because if
you use something like PVM or MPI (even over SCI) there could be
a real big loss because of syncronisation mechanisms.
With RT-Linux, you have an implicit syncronisation between tasks,
so you don't have to care about syncronisation.
Ok, you will sometimes have to force an external Syncronisation because
the CPUs don't run with the same speed, because the Oscillators on
the Mainboards are driffting a little bit ... but even this
could be solved by a single 14.1818MHz-Signal source applied to
the PLL-Signal generator on the mainboards via a shielded NF-Wire
(ok that wouldn't look fine but should work).
> If possible, I would need then an Operatin System like RTLinux?
I would call a system like the described above an
"Auto Syncronising Real-Time Cluster", and
IMHO a real-time os would be necessary ...
regards
Bernhard
--- [rtl] ---
To unsubscribe:
echo "unsubscribe rtl" | mail [EMAIL PROTECTED] OR
echo "unsubscribe rtl <Your_email>" | mail [EMAIL PROTECTED]
----
For more information on Real-Time Linux see:
http://www.rtlinux.org/~rtlinux/
