tailieunhanh - Stochastic Control Part 4

Tham khảo tài liệu 'stochastic control part 4', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | 112 Stochastic Control tn 1 N-subcube of Bellman values the node Me has to send to this node. So each node achieve the second loop of the algorithm described on figure 6 and computes P intersections of N-subcubes coordinates to get the coordinates of the P N-subcube of Bellman values it has to send. Again many of these N-subcubes are empty when the shadow regions are not too large. Figure 5 shows an example of local routing plan computed on node 1 considering the data distribution partially illustrated on figure 4. This entire routing plan computation consists in intersections of N-subcube coordinates. Finally this is a very fast integer computation run at each time step. Routing plan execution Node communications are implemented with non-blocking communications and are overlapped in order to use the maximal abilities of the interconnection network. However for large number of nodes we can get small sub-cubes of data on each node and the influence areas can reach many nodes not only direct neighbor nodes . Then the routing plan execution achieves a huge number of communications and some node interconnexion network could saturate and slow down. So we have parameterized the routing plan execution with the number of nodes that a node can attempt to contact simultaneously. This mechanism spreads the execution of the communication plan and the spreading out is controlled by two application options specified on the command line one for the optimization part and one for the simulation part. When running our benchmark see section 5 on our 256 dual-core PC cluster it has been faster not to spread these communications but on our 8192 quad-core Blue Gene P it has been really faster to spread the communications of the simulation part. Each Blue Gene node has to contact only 128 or 256 other nodes at the same time to prevent the simulation time to double. When running larger benchmarks closer to future real case experiments the size of the data and of the shadow regions