MRNet: Multicast/Reduction Network Efficient Communication for Large Scale Systems MRNet API MRNet Improves Tool Performance by Distributing Data Processing and Analyses User Interface Analysis and Control Analysis and Control MRNet Communication Processes Overlay network for efficient data multicasts and reductions. User Interface Tool Front-End MRNet Overview Tree topology for distribution of data analysis and management. Daemon0 Daemon1 Daemon2 Daemon3 Process0 Process1 Process2 Process3 Efficient multicasts for scalable control. Concurrent data channels with different aggregation schemes. Open-source licensing. Daemon1 Daemon2 Daemon3 Processn-1 Processn Process0 Process1 Process2 Process3 … Daemonn-1 Daemonn Processn-1 Processn Streams are virtual channels that bind the end-points in a communicator for group communication. Streams are created with filters that perform specified data aggregations on the data flow. Tool-specific filters can be loaded into MRNet for custom aggregations. MRNet-enabled Distributed Tool Sample Performance Improving the Performance of Paradyn’s Data Analysis Service Rate/Arrival Rate High-throughput data transfers. Daemon0 End-points, retrieved from instantiated networks, are grouped into communicators. MRNet Filters Flexible process-tree topologies. Built-in filters for elementary aggregations. Daemonn Un-enhanced Distributed Tool MRNet Features Dynamic loading of user-defined data aggregations. … Daemonn-1 Tool Back-Ends Customizable components for use in diverse tool applications. C++ based API for integration into tool front-end and daemons. Microbenchmark Results • ASCI Blue Pacific: •256x4 PowerPC 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 No performance degradation with MRNet! • 512 back-ends •~64 byte packet sizes 0 100 200 300 400 500 Number of Back-Ends 32 metrics, flat tree 16 metrics, flat tree 8 metrics, flat tree 1 metric, flat tree 32 metrics, 8-way fanout 600 • 8-way vs. Flat Tree •Startup Latency: •13.9 Speedup •Round-Trip Latency: •13.4 Speedup • Reduction Throughput •8-way: 55.23 ops/sec. •Flat tree: 4.03 ops/sec (256 back-ends). Native MRNet Filters MIN, MAX, SUM, AVG Concatenation Custom Paradyn Filters Clock synchronization. Equivalence classes for application data. Time-based data synchronization. Status Release v1.1: November 2004 Tested Platforms: Linux, Aix, Solaris Additional Information: www.paradyn.org/mrnet/