Academia.eduAcademia.edu

Figure 4 – uploaded by Andrew Lumsdaine

See full PDFdownloadDownload figure
Figure 4. BFS times on Big Red with k = 6.

Figure 4 BFS times on Big Red with k = 6.

Related Figures (10)

Algorithm 2: \’BX—Nonblocking Consensus.
Algorithm 2: \’BX—Nonblocking Consensus.
Figure 1. Visualization of different modeling parameters.
Figure 1. Visualization of different modeling parameters.
Table 1. LogGP parameters for all machines. We assume in the following that all communication algorithms are implemented optimally in the LogGP model. We measured the LogGP parameters of all test systems as described in [19, 20]. Table 1 shows the parameters for all systems.
Table 1. LogGP parameters for all machines. We assume in the following that all communication algorithms are implemented optimally in the LogGP model. We measured the LogGP parameters of all test systems as described in [19, 20]. Table 1 shows the parameters for all systems.
Figure 3. Algorithm comparison on Jaguar for k = 6.
Figure 3. Algorithm comparison on Jaguar for k = 6.
Figure 2. Algorithm comparison on Intrepid for k = 6.
Figure 2. Algorithm comparison on Intrepid for k = 6.
The communication pattern in the BFS algorithm depends on the initial graph structure and the evolution of the computation. Many real-world problems can be represented by scale-free graphs with a very small number of high-degree vertices and a sparse net- work of low-degree vertices that often form clusters. An efficient partitioning algorithm is necessary to achieve a good work distri- bution. We generated well-partitioned (balanced) test graphs, as de- Algorithm 3 shows the pseudocode for parallel level-ordered BFS. The call globalsum is a synchronizing operation that com- putes a global census function, equivalent to MPI_Allreduce with MPI_LOR as operation. This operation is used in Algorithm 3 as a global OR in order to check for termination of the algorithm. The call DSDE(R, Q, d) starts a sparse exchange of all vertices in the set R, and all received vertices are inserted into Q with their re- ceived dist. If an edge is remote, then it carries the address of the destination process as a property.
The communication pattern in the BFS algorithm depends on the initial graph structure and the evolution of the computation. Many real-world problems can be represented by scale-free graphs with a very small number of high-degree vertices and a sparse net- work of low-degree vertices that often form clusters. An efficient partitioning algorithm is necessary to achieve a good work distri- bution. We generated well-partitioned (balanced) test graphs, as de- Algorithm 3 shows the pseudocode for parallel level-ordered BFS. The call globalsum is a synchronizing operation that com- putes a global census function, equivalent to MPI_Allreduce with MPI_LOR as operation. This operation is used in Algorithm 3 as a global OR in order to check for termination of the algorithm. The call DSDE(R, Q, d) starts a sparse exchange of all vertices in the set R, and all received vertices are inserted into Q with their re- ceived dist. If an edge is remote, then it carries the address of the destination process as a property.
Figure 5. BFS times on Intrepid with k = 6.
Figure 5. BFS times on Intrepid with k = 6.
Table 3. Key properties of the used matrices and k for P = 512. ble 3 shows the number of rows, nonzero elements and k for a 512-process run. Details about the input matrices can be obtained from [30].
Table 3. Key properties of the used matrices and k for P = 512. ble 3 shows the number of rows, nonzero elements and k for a 512-process run. Details about the input matrices can be obtained from [30].
Figure 6. Number of Neighbors k for Different Applications.
Figure 6. Number of Neighbors k for Different Applications.
Related topics:
Computer ScienceAlgorithmsDistributed ComputingSparse representationSparse DataCommunication ProtocolFast AlgorithmBreadth first searchDistributed Termination
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved