Outcomes of OpenMP Hackathon: OpenMP Application Experiences with the Offloading Model (Part II)

The performance of Fortran 2008 DO CONCURRENT (DC) relative to OpenACC and OpenMP target offloading (OTO) with different compilers is studied for the GAMESS quantum chemistry application. Specifically, DC and OTO are used to offload the Fock build, which is a computational bottleneck in most quantum chemistry codes, to GPUs. The DC Fock build performance is studied on NVIDIA A100 and V100 accelerators and compared with the OTO versions compiled by the NVIDIA HPC, IBM XL, and Cray Fortran compilers. The results show that DC can speed up the Fock build by 3.0× compared with that of the OTO model. With similar offloading efforts, DC is a compelling programming model for offloading Fortran applications to GPUs.

show abstract

“…This is an example of the OpenMP kernels currently developed and implemented in GAMESS. The details of the algorithm are presented in refs and .…”

Section: Algorithmsmentioning

confidence: 99%

Enabling Fortran Standard Parallelism in GAMESS for Accelerated Quantum Chemistry Calculations

Alkan

Pham

Hammond³

et al. 2023

J. Chem. Theory Comput.

Self Cite

View full text Add to dashboard Cite

show abstract

“…As the method upon which nearly all others are based, the RHF method has been optimized and reoptimized for performance for decades by electronic structure theorists (for overviews of the topic, see refs − ). − …”

Section: Introductionmentioning

confidence: 99%

“…With the introduction of multithreading, many modern updates and developments to the RHF algorithm were made using a hybrid MPI/multithreaded scheme. − , Hybrid MPI/multithreaded algorithms place an emphasis both on the effective distribution of shell quartets of ERIs to threads in each rank in a load-balanced fashion and on the minimization of synchronization overhead introduced by multithreading while simultaneously avoiding race conditions. Two such algorithms exist in GAMESS developed by Mironov et al, one with a “ private Fock” algorithm that stores a separate copy of the Fock matrix for each thread and the other a “ shared Fock” algorithm that stores a single copy of the Fock matrix per MPI rank and utilizes mutexes to synchronize thread writes to the global Fock matrix.…”

Section: Introductionmentioning

confidence: 99%

“…A third hybrid MPI/multithreaded RHF algorithm exists in the LibCChem library associated with GAMESS. Finally, a fourth multithreaded RHF algorithm, designed for offloading to graphics processing units (GPUs), is currently under development for GAMESS. − The GTFock library − also utilizes a hybrid MPI/multithreaded scheme to parallelize the RHF Fock build process. Other notable implementations of such parallelism came in the Turbomole and NWChem packages.…”

Section: Introductionmentioning

confidence: 99%

“…OpenMP has become an extremely popular construct for use in a wide variety of HPC programs. Multiple algorithms for different methods that utilize threaded parallelism based on OpenMP exist in GAMESS; these include RHF, − RI-MP2, and RI-CC . Alongside GAMESS, the associated LibCChem library also contains multiple codes that utilize OpenMP, including RHF, RI-MP2, and CCSD(T) codes.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

A Task-Based Approach to Parallel Restricted Hartree–Fock Calculations

Poole

Vallejo

Gordon

2022

J. Chem. Theory Comput.

View full text Add to dashboard Cite

In recent years, parallelism via multithreading has become extremely important to the optimization of high-performance electronic structure theory codes. Such multithreading is generally achieved via OpenMP constructs, using a fork-join threading model to enable thread-level data parallelism within the code. An alternative approach to multithreading is task-based parallelism, which displays multiple benefits relative to fork-join thread parallelism. A novel Restricted Hartree−Fock (RHF) algorithm, utilizing task-based parallelism to achieve optimal performance, was developed and implemented into the JuliaChem electronic structure theory software package. The new RHF algorithm utilizes a unique method of shell quartet batch creation, enabling construction and distribution of fine-grained shell quartet batches in a load-balanced manner using the Julia task construct. These shell quartet batches are then distributed statically across message-passing interface (MPI) ranks and dynamically across threads within an MPI rank, requiring no explicit inter-rank or interthread synchronization to do so. Compared to the hybrid MPI/OpenMP RHF algorithm present in the GAMESS software package, the task-based algorithm demonstrates speedups of up to ∼40% for systems in the S22(3) test set of molecules, with system sizes up to ∼1000 basis functions. The JuliaChem algorithm demonstrates the viability of both the task-based parallelism model and the Julia programming language for construction of performant electronic structure theory codes targeting systems of a size of chemical interest.

show abstract