Modeling UGAL on the Dragonfly Topology
Author/Presenters
Event Type
Workshop
Accelerators
Benchmarks
Compiler Analysis and Optimization
Deep Learning
Effective Application of HPC
Energy
Exascale
GPU
I/O
Parallel Application Frameworks
Parallel Programming Languages, Libraries, Models
and Notations
Performance
Simulation
Storage
TimeMonday, November 13th9:30am -
10am
Location704-706
DescriptionThe Dragonfly topology has been proposed and deployed
as the interconnection network topology for next
generation supercomputers. Practical routing algorithms
developed for Dragonfly are based on a routing scheme
called Universal Globally Adaptive Load-balanced routing
with Global information (UGAL-G). While UGAL-G and
UGAL-based practical routing schemes have been
extensively studied, all existing results are based on
simulation or measurement. There is no theoretical
understanding of how the UGAL-based routing schemes
achieve their performance on a particular network
configuration as well as what the routing schemes
optimize for. In this work, we develop and validate
throughput models for UGAL-G on the Dragonfly topology,
and identify a robust model that is both accurate and
efficient across many Dragonfly variations. Given a
traffic pattern, the proposed models estimate the
aggregate throughput for the pattern accurately and
effectively. Our results not only provide a mechanism to
predict the communication performance for large scale
Dragonfly networks but also reveal the inner working of
UGAL-G, which furthers our understanding of the
UGAL-based routing on Dragonfly.
