Logo Logo
Hilfe
Kontakt
Switch language to English
Partial aggregation for collective communication in distributed memory machines
Partial aggregation for collective communication in distributed memory machines
High Performance Computing (HPC) systems interconnect a large number of Processing Elements (PEs) in high-bandwidth networks to simulate complex scientific problems. The increasing scale of HPC systems poses great challenges on algorithm designers. As the average distance between PEs increases, data movement across hierarchical memory subsystems introduces high latency. Minimizing latency is particularly challenging in collective communications, where many PEs may interact in complex communication patterns. Although collective communications can be optimized for network-level parallelism, occasional synchronization delays due to dependencies in the communication pattern degrade application performance. To reduce the performance impact of communication and synchronization costs, parallel algorithms are designed with sophisticated latency hiding techniques. The principle is to interleave computation with asynchronous communication, which increases the overall occupancy of compute cores. However, collective communication primitives abstract parallelism which limits the integration of latency hiding techniques. Approaches to work around these limitations either modify the algorithmic structure of application codes, or replace collective primitives with verbose low-level communication calls. While these approaches give fine-grained control for latency hiding, implementing collective communication algorithms is challenging and requires expertise knowledge about HPC network topologies. A collective communication pattern is commonly described as a Directed Acyclic Graph (DAG) where a set of PEs, represented as vertices, resolve data dependencies through communication along the edges. Our approach improves latency hiding in collective communication through partial aggregation. Based on mathematical rules of binary operations and homomorphism, we expose data parallelism in a respective DAG to overlap computation with communication. The proposed concepts are implemented and evaluated with a subset of collective primitives in the Message Passing Interface (MPI), an established communication standard in scientific computing. An experimental analysis with communication-bound microbenchmarks shows considerable performance benefits for the evaluated collective primitives. A detailed case study with a large-scale distributed sort algorithm demonstrates, how partial aggregation significantly improves performance in data-intensive scenarios. Besides better latency hiding capabilities with collective communication primitives, our approach enables further optimizations of their implementations within MPI libraries. The vast amount of asynchronous programming models, which are actively studied in the HPC community, benefit from partial aggregation in collective communication patterns. Future work can utilize partial aggregation to improve the interaction of MPI collectives with acclerator architectures, and to design more efficient communication algorithms.
Not available
Kowalewski, Roger
2021
Englisch
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Kowalewski, Roger (2021): Partial aggregation for collective communication in distributed memory machines. Dissertation, LMU München: Fakultät für Mathematik, Informatik und Statistik
[thumbnail of Kowalewski_Roger.pdf]
Vorschau
PDF
Kowalewski_Roger.pdf

1MB

Abstract

High Performance Computing (HPC) systems interconnect a large number of Processing Elements (PEs) in high-bandwidth networks to simulate complex scientific problems. The increasing scale of HPC systems poses great challenges on algorithm designers. As the average distance between PEs increases, data movement across hierarchical memory subsystems introduces high latency. Minimizing latency is particularly challenging in collective communications, where many PEs may interact in complex communication patterns. Although collective communications can be optimized for network-level parallelism, occasional synchronization delays due to dependencies in the communication pattern degrade application performance. To reduce the performance impact of communication and synchronization costs, parallel algorithms are designed with sophisticated latency hiding techniques. The principle is to interleave computation with asynchronous communication, which increases the overall occupancy of compute cores. However, collective communication primitives abstract parallelism which limits the integration of latency hiding techniques. Approaches to work around these limitations either modify the algorithmic structure of application codes, or replace collective primitives with verbose low-level communication calls. While these approaches give fine-grained control for latency hiding, implementing collective communication algorithms is challenging and requires expertise knowledge about HPC network topologies. A collective communication pattern is commonly described as a Directed Acyclic Graph (DAG) where a set of PEs, represented as vertices, resolve data dependencies through communication along the edges. Our approach improves latency hiding in collective communication through partial aggregation. Based on mathematical rules of binary operations and homomorphism, we expose data parallelism in a respective DAG to overlap computation with communication. The proposed concepts are implemented and evaluated with a subset of collective primitives in the Message Passing Interface (MPI), an established communication standard in scientific computing. An experimental analysis with communication-bound microbenchmarks shows considerable performance benefits for the evaluated collective primitives. A detailed case study with a large-scale distributed sort algorithm demonstrates, how partial aggregation significantly improves performance in data-intensive scenarios. Besides better latency hiding capabilities with collective communication primitives, our approach enables further optimizations of their implementations within MPI libraries. The vast amount of asynchronous programming models, which are actively studied in the HPC community, benefit from partial aggregation in collective communication patterns. Future work can utilize partial aggregation to improve the interaction of MPI collectives with acclerator architectures, and to design more efficient communication algorithms.