Skip to content

README.md

Latest commit

 

History

History
92 lines (70 loc) · 3.06 KB

README.md

File metadata and controls

92 lines (70 loc) · 3.06 KB
Raw

SPOOLES-MT — Multithreaded Sparse Direct Solver

POSIX-thread parallel extension to SPOOLES 2.2. Parallelises the LU factorisation and solve across multiple cores on a single node without requiring MPI. Sourced from the NguyenLabJHU/Tahoe-FEM fork, which bundles the original spoolesMT directory from the SPOOLES 2.2 distribution.

Requirements

Dependency Notes
TAHOE_SPOOLES=ON Serial SPOOLES must be built first
pthreads Standard on Linux/macOS; CMake finds via find_package(Threads)
No MPI Single-node only

Enabling

cmake -B build -DTAHOE_SPOOLES=ON -DTAHOE_SPOOLES_MT=ON
cmake --build build -j$(nproc)

How it works

SPOOLES-MT adds a parallel factorisation layer (MT.factorMT.c, MT.solveMT.c) on top of the serial SPOOLES frontal matrix data structures. The driver (LU_MT_driver) accepts a thread count at initialisation and distributes front-panel tasks across a pthread pool.

Thread-safety note on Lock.h

SPOOLES's Lock object uses compile-time selection between a no-op stub and a real pthread_mutex_t. The selector is __POSIX_THREADS__. When TAHOE_SPOOLES_MT=ON, this define is propagated to both tahoe_spooles (serial library) and tahoe_spooles_mt. This is safe: the serial LU driver always passes NO_LOCK (lockflag=0) to Lock_init, so no mutex is ever allocated in the serial code path.

Source layout

spoolesMT/src/
    MT/
        MT.factorMT.c       parallel LU factorisation kernel
        MT.solveMT.c        parallel triangular solve
        MT.mvm.c            parallel matrix-vector multiply
        MT.QRfactorMT.c     parallel QR factorisation
        MT.QRsolveMT.c      parallel QR solve
        MT.spoolesMT.h      internal header
    drivers/
        LU_MT_driver.c      high-level entry point (init/factor/solve/free)
        LU_MT_driver.h      public API
        LU_MT_driver_int.h  internal driver header
        LU_MT_driver_init.c
        LU_MT_driver_factorize.c
        LU_MT_driver_solve.c
        LU_MT_driver_free.c
    SPOOLESMT.h             top-level include

Tahoe wrapper

tahoe/src/primitives/globalmatrix/SPOOLES/
    SPOOLESMatrixT_MT.h / .cpp    pthreads LU wrapper

SolverT.cpp selects SPOOLESMatrixT_MT when both __SPOOLES_MT__ and __POSIX_THREADS__ are defined.

Compile defines

Define Set by Effect
__SPOOLES_MT__ tahoe/CMakeLists.txt (PUBLIC on libtahoe) Activates SPOOLESMatrixT_MT in SolverT.cpp
__POSIX_THREADS__ tahoe_spooles_mt (PUBLIC) Selects pthread layout in Lock.h; propagated to tahoe_spooles too

XML usage

In the solver block of your input XML, replace the default SPOOLES entry with:

<SPOOLES_MT_matrix num_threads="8" .../>

Set num_threads to the number of physical cores available. Over-subscription (more threads than cores) typically reduces performance.

Benchmark results

All 36 Google Test unit tests pass with TAHOE_SPOOLES_MT=ON. Level.0–2 benchmark results are identical to the serial build (no regressions).