Smoothed Particle Hydrodynamics with GRAPE and Parallel Virtual Machine

We have developed Remote-GRAPE (for "GRAvity PipE"), a subroutine library to be used with the special purpose computer GRAPE-3A. GRAPE-3A can efficiently calculate gravitational forces between particles and construct neighbor lists. All other calculations are performed on the host workstation (WS), which is directly connected to GRAPE. The use of GRAPE for smoothed particle hydrodynamics (GRAPE-SPH) can in principle greatly speed up the calculations on the WS. However, the current bottleneck of GRAPE-SPH is that its performance is limited by the speed of the host WS. To solve this problem, we implement Remote-GRAPE; it allows us to run applications that use GRAPE-3A hardware on significantly faster computers than the physical host WS. Thus, we can take advantage of fast computers even though they cannot physically be connected to GRAPE. The Remote-GRAPE system is implemented on Parallel Virtual Machine. The details of implementation are described.

We analyze the performance of Remote-GRAPE and obtain the following results. (1) When Remote-GRAPE is used to calculate gravitational forces only, the overhead due to the network decreases according to the number of particles (20%-40% of total time). (2) When Remote-GRAPE is used to calculate gravity and neighbor lists, the overhead due to the network does not occupy a large fraction of total time but only ~20%-30%, because the computation required on the slave machine is very large because of the properties of the GRAPE system. (3) We also compare the performance of Remote-GRAPE with the tree method. The tree method requires more time for a higher degree of clustering, while the required time for Remote-GRAPE does not much depend on it.

We then analyze the performance of GRAPE-SPH with Remote-GRAPE. The performance of Remote-GRAPE is about 4 times faster than GRAPE-SPH with usual GRAPE for our configuration. Using Remote-GRAPE, we can calculate the GRAPE part and other parts in parallel. This parallel method leads to a further speedup of our SPH code. We estimate how the performance of Remote-GRAPE depends on the configuration. We also show that the performance of Remote-GRAPE can be further improved.