Tpetra FY18 Plan

Tpetra FY18 plan

Tpetra team FY18 planning

DELIVERY: FY18 Q1; POC: Devine
Identification of performance, maintainability, quality issues; feasibility analysis; priority setting

Sparse Matrix-Matrix multiplication performance

DELIVERY: FY18 Q2; POC: Siefert
Integration of KokkosKernels SpGEMM on GPUs; performance evaluation MPI+X improvements, including algorithmic improvements ported from Epetra

Benchmark apps; platform/configuration set-up

DELIVERY: FY18 Q2 -- identification, repository, configure, build; FY18 Q4 -- automated testing; POC: Hu

• MueLu setup - use MueLu driver
• CG - use benchmark in Tpetra
• Assembly -- best practice code
• Import / Export
• Overlap - Ifpack2 Additive-Schwarz
• BlockCRS

Extended test coverage beyond Trilinos' regular testing

DELIVERY: FY18 Q2; POC: Trott

Evaluate potential gains from reducing global reductions; determine plan and priority for action

DELIVERY: Evaluation in FY18 Q2; POC: Luchini
Benchmark MueLu vs ML set-up; identify potential benefit and semantic changes

Ensure and evaluate usage of non-UVM memory for MPI communication routines

DELIVERY: FY18 Q3; POC: Hoemmen

Map / Graph / Matrix / Multivector design and performance improvements

DELIVERY: FY18 Q3 for development and deprecation; FY19 Q1 for removal of deprecated code; POC: Trott

• Includes performance improvements for Matrix / Graph construction and fill
• Includes Deprecation of Dynamic Profile (insert operation that can fail, examples of new application usage, utilities for reallocation, support for knowing max # entries but not max # entries per row)
• Includes reducing unsafe data access by returning of unmanaged Kokkos::View

Reduce number of template parameters to speed build times and reduce library size

DELIVERY: FY18 Q3 for development and deprecation; FY19 Q1 for removal of deprecated code; POC: Hoemmen

• Fix GlobalOrdinal to be int64_t
• Allow compile-time option for LocalOrdinal: int32_t or int64_t
• Determine node type from compile-time option rather than template parameter
• Includes defining offset_type to be Kokkos::size_type

Best practices documentation, User's guide

DELIVERY: FY18 Q4; POC: Fuller

As time / staffing / funding allows:

• Distributor / Import / Export design and performance
• Use KokkosKernels interfaces to on-node TPLs
• Create Tpetra::initialize, Tpetra::finalize to enable correct handling of Kokkos, MPI
• Matrix-Matrix multiplication: enable separate symbolic / numeric usage
• Remove local solves from Tpetra
• Consolidate and make consistent use of arithmetic and communication Traits
• Remove CUDA_LAUNCH_BLOCKING=1 requirement
• Overlapping computation and communication
• Separation of concerns: e.g., Comm: Teuchos::Comm, iallreduce in Tpetra
• Feature testing of ROL- and Sacado-like data types
• Remove Tpetra::Map::getIndexBase() method
• Dual view semantics
• Support for limited graph changes

January 31, 2018
SAND2018-1063 O