Number of IO MPI tasks can be varied by selecting which dimensions use parallel IO and which dimensions use Serial send to boss I/O. Thus can neck down from, say 1024 nodes = 4x4x8x8 to {1,8,32,64,128,256,1024} nodes doing the I/O. Interpolates nicely between ALL nodes write their data, a single boss per time-plane in processor space [old UKQCD fortran code did this], and a single node doing all I/O. Not sure I have the transfer sizes big enough and am not overly convinced fstream is guaranteed to not give buffer inconsistencies unless I set streambuf size to zero. Practically it has worked on 8 tasks, 2x1x2x2 writing /cloning NERSC configurations on my MacOS + OpenMPI and Clang environment. It is VERY easy to switch to pwrite at a later date, and also easy to send x-strips around from each node in order to gather bigger chunks at the syscall level. That would push us up to the circa 8x 18*4*8 == 4KB size write chunk, and by taking, say, x/y non parallel we get to 16MB contiguous chunks written in multi 4KB transactions per IOnode in 64^3 lattices for configuration I/O. I suspect this is fine for system performance.
Grid
Data parallel C++ mathematical object library
This library provides data parallel C++ container classes with internal memory layout that is transformed to map efficiently to SIMD architectures. CSHIFT facilities are provided, similar to HPF and cmfortran, and user control is given over the mapping of array indices to both MPI tasks and SIMD processing elements.
- Identically shaped arrays then be processed with perfect data parallelisation.
- Such identically shapped arrays are called conformable arrays.
The transformation is based on the observation that Cartesian array processing involves identical processing to be performed on different regions of the Cartesian array.
The library will both geometrically decompose into MPI tasks and across SIMD lanes. Local vector loops are parallelised with OpenMP pragmas.
Data parallel array operations can then be specified with a SINGLE data parallel paradigm, but optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a significant simplification for most programmers.
The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture. Presently SSE4 (128 bit) AVX, AVX2 (256 bit) and IMCI and AVX512 (512 bit) targets are supported.
These are presented as
vRealF, vRealD, vComplexF, vComplexD
internal vector data types. These may be useful in themselves for other programmers. The corresponding scalar types are named
RealF, RealD, ComplexF, ComplexD
MPI, OpenMP, and SIMD parallelism are present in the library.
You can give `configure' initial values for configuration parameters by setting variables in the command line or in the environment. Here are examples:
./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -msse4" --enable-simd=SSE4
./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -mavx" --enable-simd=AVX1
./configure CXX=clang++ CXXFLAGS="-std=c++11 -O3 -mavx2" --enable-simd=AVX2
./configure CXX=icpc CXXFLAGS="-std=c++11 -O3 -mmic" --enable-simd=AVX512 --host=none
For developers: Use reconfigure_script in the scripts/ directory to create the autotools environment