Joint Effects of Application Communication Pattern, Job Placement and Network Routing on Fat-tree Systems Peixin Qiao Illinois Institute of Technology
Mo#va#on • Simulator & Applica#ons Methodology • Fat-Tree Network • Job Placement And Rou#ng Configura#ons • Applica#on Communica#on Characteris#cs Simula#on • Performance Analysis
Motivation
Motivation
Methodology
Methodology Simulator: Co-design of Exascale Storage System (CODES) Applications: Three representative applications from Design Forward Project in DOE workload Algebraic MultiGrid Solver (AMG) Crystal Router (CrystalRouter) Geometric Multigrid V-Cycle (MultiGrid)
Fat-Tree Network Three-Tier Network Topology k-ary Fat-Tree Architecture Pruned Fat-Tree Configuration
k-ary Fat-Tree
Pruned fat-tree system Switch ports: 36 Pod: 2 Computer Nodes: 36/2 * 36/2 * 2 = 648
Job Placement 1 2 3 4 5 6 7 8 Random Placement: Each job is assigned a random set of available computer nodes. E.g.: 3 5 1 7 8 6 2 4
Job Placement 1 2 3 4 5 6 7 8 Group Placement: The computer nodes are assigned to one switch first randomly, then another switch randomly. E.g.: 3 4 7 8 5 6 1 2
Job Placement 1 2 3 4 5 6 7 8 Contiguous Placement: The computer nodes are assigned to an application consecutively. E.g .: 1 2 3 4 5 6 7 8
Network Routing Static Routing: A packet takes the static path from the source to destination. Adaptive Routing: The path a packet takes will be adaptively chosen.
Simulation
App Comm Characteristics AMG CrystalRouter Nearest-Neighbor Communication
App Comm Characteristics MultiGrid Many-to-Many Communication
Performance Analysis AMG CrystalRouter MultiGrid
AMG
AMG
Conclusion AMG CrystalRouter MultiGrid
Thanks!
Recommend
More recommend