This core project is devoted to developing efficient and stable implementations of the AMBER molecular modeling package on the Intel Paragon, specifically for use at SDSC. There are two general aspects to the proposed research: first, a fine-grained implementation of the molecular dynamics and minimization algorithihs will be implemented, followed by a more coarsegrained parallelization of free energy calculations. A new direction in the past year has involved parailelization of the particle-mesh Ewald code, along with further improvements in other areas. A major test case involving a 60,000 atom system is in progress. We have chosen to use the Message Passing Interface (MPI) programming model developed at Argonne Labs for our basic MD implementation, and have interfaced this to the ICC communication libraries provided by Intel to actually carry out the communications, which require distributed global sums and distributed global broadcasts as the most time consuming communication steps. The use of MPI points us toward an emerging standard, and allows us to use (nearly) identical parallel code on clusters of workstations and on MPP machines like the Cray T3D, IBM's SPl and SP2 machines, and the Convex Exemplar SPP. During the past year, we have also incorporated a native MPI Paragon implementation, which avoids a further software layer and allows code for the Intel Paragon machine to be identical to that for many other MPIcompliant machines. The MPI implementation is now a standard part of version 4. l of AMBER, making it available to workers throughout the world. The code development has been a collaborative effort, involving primarily John Vincent and Ken Merz (Penn State), Tom Cheatharn and Peter Kollman (UCSF), David Case (Scripps) and Jerry Greenberg and Jack Rogers (SDSC). In collaboration with Prof. J.A. McCammon, who holds appointments at SDSC and in the UCSD Department of Chemistry and Biochemistry, we have begun a large-scale simulation of protein kinase C in water. This system has nearly 60,000 atoms, and has forced us to greatly improve memory utilization in order to allow the entire program to reside in core on each node. This has been done (for the 32-Mbyte """"""""fate' nodes), and equilibration and molecular dynamics simulations are now in progress. This will provide important experience about the use of parallel machines in """"""""production"""""""" environments on very large systems. Goals for the coming year include: (l) tuning of the communication steps to achieve performance more like that expected on theoretical grounds; (2) testing and optimization of free energy calculations using this technology; (3) running additional production calculations (of DNA + salt + water, as outlined in the grant proposal) to test and illustrate the use of the Paragon in a non-benchmark environment; (4) work on adapting coarse-grained parallel algorithms for free energy calculations.
Showing the most recent 10 out of 270 publications
