Boost logo

Boost :

From: Andrey Melnikov (melnikov_at_[hidden])
Date: 2005-08-19 03:51:37


> You might want to look at the bottom of
> http://www.osl.iu.edu/research/pbgl/documentation/graph/index.html
>
"Parallel Processing Library"

http://www.osl.iu.edu/research/pbgl/documentation/parallel/

referenced there is exactly "a higher level framework for distributed
computations" I suggested you to look at.

Unfortunately, it offers very little additional abstraction over raw
MPI2 C++ API so it's advantage is doubtful.

I thought of library which offers more specialized work flow concepts.

For example, often a distributed environment is used just because there
are a lot of data to be processed, and each piece of data can be
processed completely independently or without major cooperation between
computational nodes. Examples are video rendering/processing (if it's
not I/O bound), distributed C++ compilation, SETI_at_Home/other BOINC projects.

Such tasks fit a model which I call "std::transform with a stateless
functor". MPI has abstractions for this (like MPI_Scatter/Gather) but
it's very limited, ineffective and practically useless in many
circumstances. Different node speeds, different and non-predictable
amount of resources needed to process each data unit lead to node
stalls. And MPI_Scatter requires an in-memory array instead of just an
input, a past-the-end and an output iterator.

Andrey


Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk