Collective communication routines in PVM
The collective communication routines of scatter, gather, and reduce are frequently implemented as part of the native library for parallel architectures. These operations have been implemented in PVM for use among a heterogeneous system of workstations and parallel computers forming a virtual parallel machine. In the case of the Intel Paragon machines, the PVM implementation of the reduce operation utilizes the corresponding native mode library routines whenever possible. This paper describes the implementation of these collective communication routines in PVM including the utilization of the Intel Paragon native mode operations. Performance data is also given comparing the use of the native Intel Paragon collective routines and the PVM implementation on top of these routines on a dedicated Intel Paragon. For our timing results an average latency of 109 {mu}s is incurred using PVM as compared to the native Intel global sum routine. This extra startup is independent of the size of the message being sent and the number of nodes in the group. It is demonstrated that the use of static groups is preferable in time efficiency over the use of dynamic groups.
- Research Organization:
- Oak Ridge National Lab., TN (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States)
- DOE Contract Number:
- AC05-96OR22464
- OSTI ID:
- 254958
- Report Number(s):
- ORNL/TM--13228; ON: DE96012417
- Country of Publication:
- United States
- Language:
- English
Similar Records
MPI performance evaluation and characterization using a compact application benchmark code
A users' guide to PVM (Parallel Virtual Machine)