Re: Multiple jobs on a multicore machine or cluster

[Ole Tange]

On Thu, Jan 17, 2013 at 12:58 PM, Nanditha Rao <nanditha.ec@gmail.com> wrote:

> 1. I need to run multiple jobs on a multicore (and multithreaded) machine. I
> am using the GNU Parallel utility to distribute jobs across the cores to
> speed up the task. The commands to be executed are available in a file
> called 'commands'. I use the following command to run the GNU Parallel.
>
> cat commands | parallel -j +0
>
> As per the guidance at this location- gnu parallel, this command is supposed
> to use all the cores to run this task. My machine has 2 cores and 2 threads
> per core.

I take it that you have a CPU with hyperthreading.

> The system monitor however shows 4 CPUs (CPU1 and CPU2 belong to
> core1, CPU3 and CPU4 belong to core2). Each job (simulation) takes about 20
> seconds to run on a single core. I ran 2 jobs in parallel using this GNU
> parallel utility with the command above. I observe in the system monitor

What system monitor are you using?

> that, if the 2 jobs are assigned to cpu1 and cpu2 (that is the same core),
> there is obviously no speed-up.

Why obviously? Normally I measure a speedup of 30-70% when using hyperthreading.

> They take about 40seconds to finish, which
> is about the time they would take if run sequentially. However, sometimes
> the tool distributes the 2 jobs to CPU1 and CPU3 or CPU4 (which means, 2
> jobs are assigned to 2 different cores). In this case, both jobs finish
> parallely in 20 seconds.

GNU Parallel does not do the distributing; it simply spawns jobs. The
distribution is done by your operating system.

> Now, I want to know if there is a way in which I can force the tool to run
> on different "cores" and not on different "threads" on the same core, so
> that there is appreciable speed-up. Any help is appreciated. Thanks!

If you are using GNU/Linux you can use taskset which can set a mask on
which cores a task can be scheduled on. If you want every other:
1010(bin) = 0xA. For a 128 core machine you could run:

cat commands | taskset 0xaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa parallel -j +0

> 2. Also, I want to know if there is a way to run this utility over a cluster
> of machines.. say, there are four 12-core machines in a cluster (making it a
> 48-core cluster).

cat commands | parallel -j +0 -S server1,server2,server3,server4

Please read 
http://www.gnu.org/software/parallel/man.html#example__using_remote_computers
or watch the intro videos:
https://www.youtube.com/playlist?list=PL284C9FF2488BC6D1


/Ole