Performance Evaluation of Cluster Networking and I/O Technologies

Overview

The aim of this project is to design and develop a comprehensive benchmark suite for carrying out performance evaluation of modern cluster networking and I/O technologies.

Description

Currently the benchmark suite contains the following benchmarks that can be used for performance evaluation of clusters with modern networking technologies, compute platforms and MPI libraries:

  • Evaluation of Networking Technologies only
    • Send/Recv (Latency and Bandwidth)
    • RDMA (Latency and Bandwidth)
  • Evaluation of Compute Node Architectures and MPI Libraries
    • Latency
    • Bandwidth (single-pair and multi-pair)
    • Bidirectional Bandwidth (single-pair and multi-pair)
    • Messaging Rate (single-pair and multi-pair)
  • Evaluation of Compute Node Architectures, MPI Libraries and Networking Technlogies
    • Two-sided Operations
      • Latency
      • Bandwidth (single-pair and multi-pair)
      • Bidirectional Bandwidth (single-pair and multi-pair)
      • Messaging Rate (single-pair and multi-pair)
      • Bandwidth with different Window sizes
    • One-sided Operations
      • Put: latency, bandwidth and bidirectional bandwidth
      • Get: latency, bandwidth and bidirectional bandwidth
      • Accumulate: latency
    • Collective Operation
      • Broadcast/Multicast Latency
    • Multi-threading
      • Multi-threaded latency
    • Computation and Communication Overlap
      • Send/Recv
      • RDMA Write
      • RDMA Read
  • Evaluation of Network Topology
    • Multi-hop latency
    • Displaced Ring Communication
    • Fan-out
    • Fan-in
  • Evaluation of Multi-rail Network Configurations
    • Latency
    • Bandwidth
    • Bidirectional Bandwidth

Some of these benchmarks are available for download here. The complete set of banchmarks will be available from this web page in the near future.

Sponsors

Avetec/Dice