# CASSINI: Network-Aware Job Scheduling in Machine Learning Clusters

**Authors:** Sudarsanan Rajasekaran, Manya Ghobadi, Aditya Akella  
**arXiv:** [2308.00852](https://arxiv.org/abs/2308.00852)  
**Date:** 2023-08-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2308.00852v1/

## Key points

- CASSINI is a network-aware job scheduler for machine learning clusters
- It uses a novel geometric abstraction to consider communication patterns when placing jobs on network links
- An affinity graph is used to adjust communication phases and ensure interleaving of jobs on the same link
- Experiments with 13 ML models showed significant improvements in completion time and packet reduction compared to state-of-the-art schedulers
- The text also discusses parallelism techniques used in DNN training systems, such as model parallelism and tensor parallelism
- Communication patterns during different phases of parallelization approaches are discussed
- Overall, CASSINI optimizes job scheduling based on network-awareness and improves efficiency in machine learning cluster management.

## Summary

CASSINI is a network-aware job scheduler designed specifically for machine learning (ML) clusters. It introduces a novel geometric abstraction that takes into consideration the communication patterns of different jobs when placing them on network links. This is achieved through the use of an affinity graph, which identifies time-shift values to adjust the communication phases of a subset of jobs. By doing so, CASSINI ensures that the communication patterns of jobs sharing the same network link are interleaved with each other. In order to evaluate its performance, experiments were conducted using 13 common ML models on a 24-server testbed. The results showed that compared to state-of-the-art ML schedulers, CASSINI significantly improves both the average and tail completion time of jobs by up to 1.6 times and 2.5 times, respectively. Additionally, CASSINI was found to reduce the number of ECN marked packets in the cluster by up to 33 times. The expanded context provides further insights into parallelism techniques used in DNN training systems such as model parallelism, tensor parallelism, and hybrid data/pipeline/tensor parallelism. These techniques involve partitioning models horizontally or vertically across multiple servers and distributing tensors or data across workers. The communication patterns during different phases of these parallelization approaches are also discussed. Overall, CASSINI's ability to optimize job scheduling based on network-awareness and its demonstrated improvements in completion time and packet reduction make it a valuable tool for efficient machine learning cluster management.