Balancing job performance with system performance via locality-aware scheduling on torus-connected systems

Abstract

Torus-connected network is widely used in modern supercomputers due to its linear per node cost scaling and its competitive overall performance. Job scheduling system plays a critical role for the efficient use of supercomputers. As supercomputers continue growing in size, a fundamental problem arises: how to effectively balance job performance with system performance on torus-connected machines? In this work, we will present a new scheduling design named window-based locality-aware scheduling. Our design contains three novel features. First, rather than oneby-one job scheduling, our design takes a "window" of jobs, i.e. multiple jobs, into consideration for job prioritizing and resource allocation. Second, our design maintains a list of slots to preserve node contiguity information for resource allocation. Finally, we formulate our scheduling decision making into a 0-1 Multiple Knapsack Problem and present two algorithms to solve the problem. A series of trace-based simulations using job logs collected from production supercomputers indicate that this new scheduling design has real potentials and can effectively balance job performance and system performance.

Topology-aware job mapping

“…Similar strategies have been recently incorporated within SLURM with 6 (or without 7 ) the use of ALPS. Another interesting work (Yang et al, 2014) adapted only for torus topology and presented a window-based locality-aware job-scheduling strategy that tries to optimize job and system performance in the same time. Its goal is to preserve node contiguity by considering multiple jobs for scheduling while making use of the 0-1 multiple knapsack problem for resource allocation.…”

A convergence of key‐value storage systems from clouds to supercomputers

“…Newly inserted or modified key‐value data will be replicated asynchronously to secondary replicas that have closer hashed location. By communicating only with near neighbors, this approach ensures that replicas only consume less network resources when we succeed in implementing the topology‐aware and locality‐aware protocols (similar approach can be found in ). Despite the lack of topology‐aware in the current ZHT, the asynchronous replication only adds relatively small overhead when adding more replicas at modest scales (up to 4K cores).…”

DRAS-CQSim: A reinforcement learning based framework for HPC cluster scheduling

“…The underlying CQSim scheduling simulator has been successfully supporting a number of projects in this field over a decade [8,[19][20][21][22][23][24][25][26][27][28][29][30]. CQSim provides a unified platform to evaluate the performance of various methods with minimal overheads.…”