Anton A. Zabreyko

Anton A. Zabreyko

Auto-generated profile, based on publicly available information

Anton A. Zabreyko

MIT, Cambridge, MA, USA

Research Focus & Metrics

Representative Publications

View All Publications
18 November 2024
Verified
Conference Paper
Open access
SRSudarsanan Rajasekaran
Sudarsanan Rajasekaran
SNSanjoli Narang
Sanjoli Narang
AZAnton A. Zabreyko
Anton A. Zabreyko
MGManya Ghobadi
Manya Ghobadi
This paper argues that congestion control protocols in machine learning datacenters sit at a sweet spot between centralized and distributed flow scheduling solutions. We present MLTCP, a technique to augment today's congestion control algorithms to approximate an interleaved centralized flow schedule. At the heart of MLTCP lies a straight-forward principle based on a key conceptual insight: by scaling the congestion window size (or sending rate) based on the number of bytes sent at each iteration, MLTCP flows eventually converge into a schedule that reduces network contention. We demonstrate that MLTCP uses a gradient descent trend with a step taken at every training (or fine-tuning) iteration towards reducing network congestion among competing jobs.

11 August 2026
Verified
Conference Paper
Open access
AZAnton A. Zabreyko
Anton A. Zabreyko
WWWeiyang Wang
Weiyang Wang
MGManya Ghobadi
Manya Ghobadi

19 November 2026 in Proceedings of the ACM on Networking
Verified
Journal Article
Open access
CLChenning Li
Chenning Li
AZAnton A. Zabreyko
Anton A. Zabreyko
OCOm Chabra
Om Chabra
ANArash Nasr-Esfahany
Arash Nasr-Esfahany
KZKevin Zhao
Kevin Zhao
TAThomas Anderson
Thomas Anderson
Distributed-application simulators need to model the completion times of individual network transfers because later computation or communication may depend on the timing of earlier completions. Packet-level simulators can model detailed queue and protocol dynamics but are too slow for application modeling. Flow-level models are scalable but inaccurate; they ignore queue buildup and transient contention effects. Existing learned network models are fast but typically are limited to preset traces rather than acting in response to dynamic application behavior. We present m4, a learned flow-level network backend that maintains state across flow arrivals and completions and returns individual completion callbacks. m4 represents the network as a stateful flow-link bipartite graph and uses dense supervision from intermediate signals to improve accuracy. We first train m4 on packet-simulator labels and evaluate on standalone simulation, where it reduces mean error by up to 57.3% versus a flow-level simulator. We then integrate m4 into SimAI, where it achieves 10.9% mean error (vs. 23.3% for the flow-level backend) while delivering a median 40.1x speedup over the packet-level backend. On an RDMA testbed, we fine-tune the checkpoint on 4-host traces and evaluate on 12-host hardware, with median application completion-time error of 15.1% vs. 31.0% for a calibrated packet-level backend and 88.4% for the flow-level backend.