[gmx-users] run g_cover in parallel

[Mark Abraham]

Tsjerk Wassenaar wrote:
> Hi,
> 
>>> I wonder if there a way to run g_cover in paralel in order to make
>>> things run faster?
>> No. Obviously you can use -dt to reduce the number of frames you analyze.
> 
> In most cases it's not the reading of frames, but the diagonalization
> of the covariance matrix that consumes most of the time.

True, such an algorithm for N atoms over F frames is either something 
like O(F) or O(N^3) according to whether the I/O-and-compute phase or 
the diagonalization phase dominates. Once could also reduce N by 
coarse-graining the covariance matrix.

> It might be
> interesting to have it MPI enabled, especially since systems are
> getting bigger and bigger.

Someone would have to identify a candidate MPI diagonalization algorithm 
that can be readily added. That doesn't sound like a easy algorithm to 
write with numerical stability :-)

ScaLAPACK http://www.netlib.org/scalapack/html/scalapack_single.html 
might work - how about http://www.netlib.org/scalapack/single/pssyevd.f 
? It would add further dependency to the build process, unless we 
incorporated it as with BLAS and LAPACK.

Mark