Barnabas Debreczeni wrote:
>I am using PGAPack as a GA library, and it uses MPI to parallelize
>optimization runs. This is how I got to Open MPI.
Let me see if I understand the underlying premise. You want to
parallelize, but there are some large shared tables. There are many
different parallelization models. E.g., there are certainly
*shared-memory* parallel programming models such as OpenMP (which is
totally different from Open MPI, despite the similar names). But you
are using MPI (which doesn't really do shared memory) since you're
trying to leverage PGAPack, which is nice for handling genetic
algorithms but basically forces you to use MPI. (I suspect most GA
algorithms map reasonably well to MPI. Your interest in shared tables
gives your situation a different twist.)
>My problem is, I'd like to share that 2 GB table (computed once at the
>beginning, and is read-only after) between processes so I don't have
>to use up 16 gigs of memory.
>How do you share data between processes locally?
Are there shared-memory parallel GA packages that might make more sense
to use here than PGAPack?
If you want to stick with PGAPack/MPI, then you can set up shared memory
among MPI processes by going outside of MPI. (You could use MPI calls
to share data, including MPI_Get routines, but I'm guessing it's best
just to add non-MPI code to do the sharing.) You can for example create
a file that each process "mmap"s into its address space. There are also
System V shared-memory calls like shmget/shmat/shmdt that allow you to
share memory among processes.
The main point: while MPI allows communication (and therefore "data
sharing") among processes, you might be better off with non-MPI
mechanisms here like mmap or SysV shared memory.
>Later I will need to use other hosts too in the calculation. Will the
>slaves on other hosts need to calculate their own tables on go on from
>there and share them locally, or can I share these tables on the
>master host with them?
I think this is a performance-vs-memory question. If your interconnect
is fast enough or your performance requirement low enough and your
memory constraints severe enough, then you can share common data among
all your nodes. You'd probably want to use MPI calls to do so...
possibly using one-sided MPI_Get routines depending on what sort of
cluster you're running on.
But, if your interconnect is not fast enough or your performance
requirement high enough or your memory constraint not too severe, then
just share within each node. And, I could imagine you might have enough
memory per node (a few Gbytes) that this will be your scenario. So,
just replicate your mmap/SysV solution on each node.
Short answer: you probably want to use non-MPI mechanisms to effect
your shared memory.
Most importantly, when your algorithm is successfully implemented and
deployed and you're making millions of dollars, please remember us!