MPI vs Threads,
MPI can be as quick as threads. Only if the
comunication between processing units (threads, or mpi
process), is 0.
Speed: disk > mpi > shared memory > cache
Bare MCGO (no UCT, nothing) can be 99.999% efficient.
Because It doesn't need inter-process communications.
A "master" process will spawn 4000 process. Every
process will run for a couple of minutes, and each one
will return the results to master. There is only two
basic communications: spawning -sending parameters-,
and finishing -gathering results and accounting-.
Never a process will wait for the result of another
process in order to continue. (very important.)
I don't remember exactly how its described, but the
"efficiency" of a parallel algorithm Its measured by
the penalty imposed in the paralelisation, compared
with pure processing power.
I.E.
1 proc, 10 minutes,10 solutions. (uni proc)
10 procs, 2 minute, 10 solutions. (multi proc)
In this example: 10/10 = 1 solution/minute vs 10/2 = 5
solutions per minute. So, In this example, the
"parallelisation penalty" is 50%. (quite common.)...
The thing is, you can run zillons processors, so you
don't care "at all" about wasting 50% of your
processing power. A 10 or 20% penalisation is quite
good efficiency... And it depends on the algorithm you
are implementing. Its inter-process communications
needs. Its "sequential" parts.
As soon as I get some spare free time, I will
implemente a parallel UCT MCGO. It will utilise a
"lazy update" transposition table. You will be able to
look at it in my opensource project. Its all in my
head.
Cheers
Eduardo
--- Darren Cook <[EMAIL PROTECTED]> escribió:
> program? Can MPI be as quick as threads on a 2- or
> 4-core single
> machine? [2] Playouts have practically no memory
> demands.
__________________________________________________
Preguntá. Respondé. Descubrí.
Todo lo que querías saber, y lo que ni imaginabas,
está en Yahoo! Respuestas (Beta).
¡Probalo ya!
http://www.yahoo.com.ar/respuestas
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
