Performance improvements¶
Instant-submission mode enabled by default when building with AdaptiveCpp¶
In GROMACS 2024, one had to manually enable the instant-submission mode
of AdaptiveCpp when building GROMACS
(-DSYCL_CXX_FLAGS_EXTRA=-DHIPSYCL_ALLOW_INSTANT_SUBMISSION=1).
Now it is enabled by default, improving performance by up to 20%
when running on GPU and slightly reducing the CPU usage when using
SYCL/AdaptiveCpp backend.
More OpenMP parallelization¶
OpenMP multithreading is now used for position restraints, GPU LINCS initialization, domain decomposition state vector sorting and a few other places. This is particularly relevant for runs where a fast GPU is bottlenecked by DD and other CPU tasks.
GPU buffer operations enabled by default¶
This is generally faster on modern GPUs, but, for very small systems, this can worsen performance by a few percent.
In this case, the old behavior can be restored by setting GMX_GPU_DISABLE_BUFFER_OPS=1 environment variable.
NVSHMEM enabled PP Halo Exchange of Forces and Coordinates¶
Introduced pack/unpack kernels fused with data transfers enabled via NVSHMEM for particle-particle (PP) halo exchange,
optimizing performance for all simulation input types. This feature enhances scalability
by leveraging NVSHMEM’s low-overhead, asynchronous communication capabilities and is supported on NVIDIA Volta and newer
architectures. Using this functionality requires NVSHMEM to be enabled at build-time (-DGMX_NVSHMEM=ON) as well as explicitly
requested at runtime with the GMX_ENABLE_NVSHMEM=1 environment variable. It should be noted that this is an early implementation,
and performance improvements for this feature are planned for a future release.