[Bf-committers] Sunday meeting minutes, 25th feb 2007

[Timothy Baldridge]

After looking into it, I agree, it looks like OpenMP is here to stay
and will soon replace pthreads as the industry standard for multi
threading. Also, I noticed something really interesting, it seems that
Intel has a Cluster OpenMP compiler that allows OpenMP to be used on a
cluster. This is an interesting concept, if only it was open
source....

But still even Cluster OpenMP would not support heterogeneous
clusters. Which is something a custom designed cluster API in blender
would allow (blender already handles all the endian issues).

The way I see it is, there are three big problems with going
multiprocess  (is that even a word?) with a program. 1)
Endian/Platform issues 2) Normal threading issues (problems with
shared memory, etc.), and 3) Serialization for streaming objects
through the network. But, with multi-threaded Blender we already have
most of those issues ironed out. Is there any reason for not having
Blender MPI enabled? Especially if there is a way to turn it on and of
at startup.

We're talking huge performance boosts. Not to mention that it would be
a big publicity boost, and something awesome to show off at Siggraph.
Just imagine showing people a Mac, a Linux box, a Solaris machine, and
a Win32 workstation all networked and all calculating a fluid
simulation.

Okay, let the arguments about my sanity begin. ;-)

Timothy

-- 
re·cur·sive      /rɪˈkɜrsɪv/: Adj. See Recursive.