arXiv subjects

Eashan Gupta

Publications and source records attributed to Eashan Gupta.

3 recordsLinked to original sources

arXiv2026

Tuning Collective Patterns to Alleviate Congestion in Shared AI Clusters

Distributed AI training involves recurring rounds of data exchange between multiple pairs of GPU nodes. Slowdown in even one flow due to congestion can cause the entire communication round to slowdown. Current approaches for evading congestion in AI clusters assume global control over the entire workload (e.g. coordinating the schedule of all jobs) or assume infrastructural support (e.g. adaptive routing in switches). They are thus ill-suited in a shared cloud setting where AI jobs belonging to one user can face external congestion from other users' jobs or background traffic beyond its own control. In this paper, we build a system, REACT, that tunes the recurring pattern of data exchange between GPU nodes (known as communication collectives) in response to congestion. REACT works at the application (communication library) layer, where it detects congestion at runtime using readily available flow stats, and tunes the collective pattern to alleviate congestion - changing the set of incident flows while retaining the semantics of information exchange (e.g. selecting which node aggregates data in an AllReduce tree). REACT requires no explicit support from the underlying network infrastructure and can be unilaterally deployed by individual users in a shared cloud setting. We prototype REACT as a shim layer over NCCL, and evaluate it on a shared academic GPU cluster - enabling REACT improves communication performance (algorithm bandwidth) by 13%-38% under network congestion. Our simulations across a range of congestion scenarios further reveal up to 75% performance improvement, highlighting the effectiveness of our approach.

cs.NI↗

arXiv2023

Upper Bounds for All and Max-gain Policy Iteration Algorithms on Deterministic MDPs

Policy Iteration (PI) is a widely used family of algorithms to compute optimal policies for Markov Decision Problems (MDPs). We derive upper bounds on the running time of PI on Deterministic MDPs (DMDPs): the class of MDPs in which every state-action pair has a unique next state. Our results include a non-trivial upper bound that applies to the entire family of PI algorithms; another to all "max-gain" switching variants; and affirmation that a conjecture regarding Howard's PI on MDPs is true for DMDPs. Our analysis is based on certain graph-theoretic results, which may be of independent interest.

cs.DM↗

arXiv2023

DBO: Response Time Fairness for Cloud-Hosted Financial Exchanges

In this paper, we consider the problem of hosting financial exchanges in the cloud. Financial exchanges require predictable, equal latency to all market participants to ensure fairness for various tasks, such as high speed trading. However, it is extremely difficult to ensure equal latency to all market participants in existing cloud deployments, because of various reasons, such as congestion, and unequal network paths. In this paper, we address the unfairness that stems from lack of determinism in cloud networks. We argue that predictable or bounded latency is not necessary to achieve fairness. Inspired by the use of logical clocks in distributed systems, we present Delivery Based Ordering (DBO), a new approach that ensures fairness by instead correcting for differences in latency to the participants. We evaluate DBO both in our hardware test bed and in a public cloud deployment and demonstrate that it is feasible to achieve guaranteed fairness and sub-100 microsecond latency while operating at high transaction rates.

cs.NI↗