The original proposal submitted to the dCSE panel requested 12 months effort. The original objectives for this project were as follows:
Work package 1: Familiarization and porting to HECToR. (Effort: 1 month, Full time)
By mid-May Dr Michael Lysaght will be familiar with running a standard, non-parallel run on HECToR (the water test currently available in the CCPForge site will be appropriate for this)
Work package 2: Parallelisation of the Hamiltonian construction. (Effort: 3.5 months, Full time)
Work package 3: Optimization of a sparse-matrix vector multiplication mixed OpenMP-MPI approach. (Effort: 3 months, Full time)
Work package 4: Implementation and optimization of PARPACK for the diagonalization step. (Effort: 4 months, Full time)
Work package 5: Final Report. (Effort: 0.5 months, Full time)
To be produced in html and .pdf format ready for publication on the HECToR website. Status update on code being released to users.
For the milestones in Work Packages 2, 3 and 4, the code developments will be demonstrated on three different tests that will cover several different characteristics of the calculations we hope to run in the future:
All these systems have been extensively studied at the OU and UCL, so a variety of different models (larger and smaller Hamiltonians) can be provided if needed. Each test case will include a Hamiltonian with 0.5 million configurations (a 500,000 x 500,000 matrix). Preliminary tests on HECToR show that the sparse matrix vector multiplication for a Hamiltonian of this size can be performed within a single Phase 2b XE 6, 24 core node. However, it will be fundamental for milestone 3 that the tests are run on two Phase 2b equivalent nodes. It is also expected that the Hamiltonian construction (milestone 2) will need to run on more than one node, so efficiency will be tested when running on two nodes. The test for milestone 4 is expected to run on 24 cores (single node).