Qlua
Lua 5.1 with lattice QCD built in.
Qlua is an interpreter for lattice QCD calculations. It is Lua 5.1 — the whole
language, unchanged — with the lattice added to it: the QDP/C field types and their
algebra, the file formats a calculation reads and writes, and a set of Level III
inverters, all reachable as ordinary Lua values. A gauge field is a variable, a
propagator is a variable, and the arithmetic between them is written the way it is
written on paper.
A calculation is therefore a script rather than a program to be compiled, and the
part of it that is physics stays the size of the physics. The parallelism is not in
the script: Qlua runs under MPI across the whole machine, and a field operation is
collective over the ranks that hold the lattice. What a script says is what is
computed, not how it is distributed.
It is built for the machines lattice calculations actually run on: the build
system ships targets for Wombat at MIT and for Perlmutter at NERSC, on CPUs and on
GPUs, and writing a target for another cluster is a short file. It also builds and
runs on a laptop, which is the way to learn it and to get a script right before it
goes to a cluster. The documentation covers
installation, a set of tutorials, and a reference manual for every name the
interpreter answers to.
The USQCD software stack
Qlua does not reimplement the lattice. It is an application package on top of the
USQCD SciDAC software stack — a suite
of modules, developed under SciDAC and layered by level, that lets a lattice QCD
computation run with high performance across a variety of architectures. Qlua is
listed there among the applications built on it, beside Chroma, CPS, FUEL and MILC.
These are the modules a Qlua build draws on:
| Level | Module | What it provides |
|---|
| 1 |
QMP |
the message-passing layer the rest of the stack communicates through |
| 1 |
QLA |
the per-site linear algebra — SU(3) matrices, spinors, and the operations on them |
| 2 |
QDP/C |
lattice-wide data-parallel fields, which are the field types Qlua exposes |
| 2 |
QIO |
record-oriented I/O in the community file formats |
| 3 |
QOPQDP |
optimized inverters, including the QOPQDP multigrid solver |
| 3 |
MDWF |
the Möbius domain wall fermion solver |
| 3 |
QUDA |
the GPU inverters, in builds configured for them |
Alongside these a build takes the clover and twisted-mass solvers, Lua itself, and the
general-purpose libraries — GSL, LAPACK, ARPACK, HDF5 — that a calculation needs.
The whole stack is built from one tarball, which is on the
downloads page; the
installation page says how. More about the
collaboration is at usqcd.org.
|
