APBS Group Meeting, June 2014

###APBS release discussion

1.4.1 release targets:

• osx – mavericks
• linux – boltz: redhat 6.5 dynamic and static 64
• windows – 64 bit

Done: built and tested on everything, code is frozen,

Todo: turn on optimization flags (O2) for all platforms and test again. Then generate install images for sourceforge.

Openmp – does that change the test suite? If so, make sure the flag is working so the user is informed that it’s working.

###1.5 release targets:

same as 1.4.1

no need to work with package developers (yast, rpms, etc)

###Libraries to include

Main libraries: zlib, readline (dependency from maloc), mpi, openmp, fetk

OpenMP – will work on some platforms – is there a command line option for this?

MPI – no binaries compiled with MPI (causes problems). Users who want MPI need to build their own source code.

Worst case: errors at runtime looking for libraries, Best case: different results

Todo: Need to document this for users on the web site!

###FETK

Best for big systems, few people use it, slow…

Need to replace maloc (hardware extraction library)

Goal – build APBS without fetk. Users who want to use it need to download it and build it on their own. Remove maloc from APBS (need portable replacement for fprintf – different for windows, memory management); For now- keep current internal maloc library and make incremental changes. During build, find fetk, if not found, use internal maloc.

1.5 target - make robust: Geometric flow and boundary element code (includes a lot of fortran – not used right now) needs to be added in to distinguish us from other solvers.

iAPBS – do not need to test against it.

mAPBS – all matlab implementation for APBS

###PDB2PKA

Issues

• checkpointing files that get read in regardless of how pdb2pka is run (python pickle files) – solution: always restart from the beginning unless a specific flag has been set.
• pdb2pqr – modified to use titration curve to make pqr file – (easiest part)
• handling all the output files from pdb2pka (use a prefix flag)
• apbs 1.4.1 library swig’d

###GPU updates

SOR – fastest solver for this problem

GSRB – easy to parallelize

Time spent – input to CPU solver is formatted for GSRB for CG

Bottlenecks: 1) Problem setup – translating protein geometry into coefficients (fillco routines, molac) 2) solvers

AMGAX – algebraic multigrid – closest to GSRB (in cusp library) – better for large system

CG or SOR better for smaller systems

Poisson solver humholds – basic eliptic pde solver

Replace functionality in PMG directory – all candidates for GPU replacements (SOR, CG, multigrid)

GPU accelerated linear algebra library – “cula”

CG needs a preconditioner – taking 1000s of iterations to converge