Congestion control in machine learning clusters

Rajasekaran, Sudarsanan; Ghobadi, Manya; Kumar, G. Santhosh

doi:10.1145/3563766.3564115

Cited by 15 publications

(1 citation statement)

References 37 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Saba requires no modification to applications and no prior knowledge of flow size. Rajasekaran et al observe that unfairly (unequally) sharing the network among ML jobs could lead to shorter training time due to the on-off pattern of DNN training [49]. In contrast, Saba does not make any assumption about the specific pattern of communication and computation, and proposes a general methodology to allocate bandwidth across a heterogeneous set of workloads.…”

Section: Related Workmentioning

confidence: 99%

Saba: Rethinking Datacenter Network Allocation from Application's Perspective

Katebzadeh

Costa

Grot

2023

Proceedings of the Eighteenth European Conference on Computer Systems

View full text Add to dashboard Cite

Today's datacenter workloads increasingly comprise distributed data-intensive applications, including data analytics, graph processing, and machine-learning training. These applications are bandwidth-hungry and often congest the datacenter network, resulting in poor network performance, which hurts application completion time. Efforts made to address this problem generally aim to achieve max-min fairness at the flow or application level. We observe that splitting the bandwidth equally among workloads is sub-optimal for aggregate application-level performance because various workloads exhibit different sensitivity to network bandwidth: for some workloads, even a small reduction in the available bandwidth yields a significant increase in completion time; for others, the completion time is largely insensitive to the available bandwidth.Building on this insight, we propose Saba, an applicationaware bandwidth allocation framework that distributes network bandwidth based on application-level sensitivity. Saba combines ahead-of-time application profiling to determine bandwidth sensitivity with runtime bandwidth allocation using lightweight software support with no modifications to network hardware or protocols. Experiments with a 32server hardware testbed show that Saba improves average completion time by 1.88× (and by 1.27× in a simulated 1,944server cluster).

show abstract

Section: Related Workmentioning

confidence: 99%