arXiv subjects

Vamsi Addanki

Publications and source records attributed to Vamsi Addanki.

18 recordsLinked to original sources

arXiv2026

Birkhoff Decompositions and Photonic Interconnects Wait! Don't Forget the Compute!

The growing demand for efficient communication in distributed training and inference has sparked significant interest in reconfigurable photonic interconnects across both academia and industry. Mixture-of-Experts (MoE) models, with their highly skewed communication patterns, present a natural opportunity for such circuit-switched fabrics. However, existing approaches largely optimize communication in isolation, overlooking the interaction between communication and the expert computation that follows. In this paper, we revisit circuit scheduling for all-to-all communication in MoE execution. We show that the dispatch--compute--combine structure fundamentally challenges classical scheduling techniques such as Birkhoff--von Neumann (BvN) decomposition. First, MoE communication matrices are rarely doubly stochastic, introducing significant scheduling bubbles in BvN-based schedules. Second, while decomposition enables communication--compute overlap, the excessive number of matchings produced by BvN fragments execution into small batches, leading to severe compute inefficiencies due to fixed execution overheads. Motivated by these observations, we explore a simple greedy max-weight decomposition strategy that bounds the number of matchings while preserving large batch sizes per matching. Despite its simplicity, the approach significantly improves overlap efficiency, reduces compute overheads, and approaches the performance of an ideal congestion-free all-to-all.

cs.NI↗

arXiv2026

BShare: Packet Queueing Delay-Driven Buffer Sharing for Datacenter Switches

Modern datacenter switches share packet buffers across ports to boost overall throughput and reduce packet loss. However, as buffer availability per-port-per-bandwidth unit continues to decrease, existing buffer-sharing strategies face increasing performance challenges. Recent efforts have attempted to integrate Buffer Management (BM) with Active Queue Management (AQM) to harness the advantages of both BM and AQM approaches to improve performance. While these hybrid solutions show promise, their complexity of dynamically calculating multiple factors for integration hinders generalization and efficiency. This paper presents BShare, a simple buffer sharing mechanism that uses packet queueing delay. BShare requires only a single operator-configurable parameter. Our simulation results show that BSHARE improves the flow completion time (FCT) performance of advanced transport protocols, such as PowerTCP, by up to 45.07% compared to ABM, particularly under burst-heavy datacenter workloads.

cs.NI↗

arXiv2026

Trivance: Latency-Optimal AllReduce by Shortcutting Multiport Networks

AllReduce is a fundamental collective communication operation in distributed computing and a key performance bottleneck for large-scale training and inference. Its completion time is determined by the number of communication steps, which dominate latency-sensitive workloads, and the communication distance affecting both latency- and bandwidth-bound regimes. Direct-connect topologies, such as Google's TPUv4 tori, are particularly prone to large communication distances due to limited bisection bandwidth. In this paper, we present Trivance, a novel AllReduce algorithm that completes within $\log_3 n$ steps - a 50% improvement in comparison to Swing and Recursive Doubling, while reducing congestion compared to Bruck's algorithm by a factor of three and preserving bandwidth-optimality. Trivance exploits both transmission ports of a bidirectional ring within each step to triple the communication distance along both directions simultaneously. By performing joint reductions, Trivance improves both the number of steps and network congestion. We further show that Trivance extends naturally to multidimensional torus networks, retaining its latency advantage while achieving performance comparable to bandwidth-optimal algorithms for large AllReduce sizes. Our packet-level SST simulation shows that Trivance improves state-of-the-art approaches by 5-30% for AllReduce sizes up to 8 MiB, in high-bandwidth settings up to 32 MiB and for 3D tori up to 128 MiB. Throughout the evaluation, Trivance remains the best-performing latency-optimal algorithm.

cs.DC↗

arXiv2026

To Reconfigure or Not to Reconfigure: Optimizing All-to-All Collectives in Circuit-Switched Photonic Interconnects

All-to-all collective communication is a core primitive in distributed machine learning and high-performance computing. At the server scale, the communication demands of these workloads are increasingly outstripping the bandwidth and energy limits of electrical interconnects, driving a growing interest in photonic interconnects. However, leveraging these interconnects for all-to-all communication is nontrivial. The core challenge lies in jointly optimizing a sequence of topologies and flow schedules, reconfiguring only when the transmission savings from traversing shorter paths outweigh the reconfiguration cost. Yet the search space of this joint optimization is enormous. Existing work sidesteps this challenge by making unrealistic assumptions on reconfiguration costs so that it is never or always worthwhile to reconfigure. In this paper, we show that any candidate sequence of topologies and flow schedules can be expressed as a sum of adjacency matrices and their powers. This abstraction captures the entire solution space and yields a lower bound on all-to-all completion time. Building on this formulation, we identify a family of topology sequences with strong symmetry and high expansion that admits bandwidth-efficient schedules, which our algorithm constructs with low computational overhead. Together, these insights allow us to efficiently construct near-optimal solutions, effectively avoiding enumeration of the combinatorial design space. Evaluation shows that our approach reduces all-to-all completion time by up to 44% on average across a wide range of network parameters, message sizes and workload types.

cs.NI↗

arXiv2026

Harvest: Adaptive Photonic Switching Schedules for Collective Communication in Scale-up Domains

As chip-to-chip silicon photonics gain traction for their bandwidth and energy efficiency, their circuit-switched nature raises a fundamental question for collective communication: when and how should the interconnect be reconfigured to realize these benefits? Establishing direct optical paths can reduce congestion and propagation delay, but each reconfiguration incurs non-negligible overhead, making naive per-step reconfiguration impractical. We present Harvest, a systematic approach for synthesizing topology reconfiguration schedules that minimize collective completion time in photonic interconnects. Given a collective communication algorithm and its fixed communication schedule, Harvest determines how the interconnect should evolve over the course of the collective, explicitly balancing reconfiguration delay against congestion and propagation delay. We reduce the synthesis problem into a dynamic program with an underlying topology optimization subproblem and show that the approach applies to arbitrary collective communication algorithms. Furthermore, we exploit the algorithmic structure of a well-known AllReduce algorithm (Recursive Doubling) to synthesize optimal reconfiguration schedules without using any optimizers. By parameterizing the formulation using reconfiguration delay, Harvest naturally adapts to various photonic technologies. Using packet-level and flow-level evaluations, as well as hardware emulation on commercial GPUs, we show that the schedules synthesized by Harvest significantly reduce collective completion time across multiple collective algorithms compared to static interconnects and reconfigure-every-step baselines.

cs.NI↗

arXiv2025

The Harmonic Policy for Online Buffer Sharing is (2 + ln n)-Competitive: A Simple Proof

The problem of online buffer sharing is expressed as follows. A switch with $n$ output ports receives a stream of incoming packets. When an incoming packet is accepted by the switch, it is stored in a shared buffer of capacity $B$ common to all packets and awaits its transmission through its corresponding output port determined by its destination. Each output port transmits one packet per time unit. The problem is to find an algorithm for the switch to accept or reject a packet upon its arrival in order to maximize the total number of transmitted packets. Building on the work of Kesselman et al. (STOC 2001) on split buffer sharing, Kesselman and Mansour (TCS 2004) considered the problem of online buffer sharing which models most deployed internet switches. In their work, they presented the Harmonic policy and proved that it is $(2 + \ln n)$-competitive, which is the best known competitive ratio for this problem. The Harmonic policy unfortunately saw less practical relevance as it performs $n$ threshold checks per packets which is deemed costly in practice, especially on network switches processing multiple terabits of packets per second. While the Harmonic policy is elegant, the original proof is also rather complex and involves a lengthy matching routine along with multiple intermediary results. This note presents a simplified Harmonic policy, both in terms of implementation and proof. First, we show that the Harmonic policy can be implemented with a constant number of threshold checks per packet, matching the widely deployed \emph{Dynamic Threshold} policy. Second, we present a simple proof that shows the Harmonic policy is $(2 + \ln n)$-competitive. In contrast to the original proof, the current proof is direct and relies on a 3-partitioning of the packets.

cs.DS↗

arXiv2025

When Light Bends to the Collective Will: A Theory and Vision for Adaptive Photonic Scale-up Domains

As chip-to-chip silicon photonics gain traction for their bandwidth and energy efficiency, collective communication has emerged as a critical bottleneck in scale-up systems. Programmable photonic interconnects offer a promising path forward: by dynamically reconfiguring the fabric, they can establish direct, high-bandwidth optical paths between communicating endpoints -- \emph{synchronously and guided by the structure of collective operations} (e.g., AllReduce). However, realizing this vision -- \emph{when light bends to the collective will} -- requires navigating a fundamental trade-off between reconfiguration delay and the performance gains of adaptive topologies. In this paper, we present a simple theoretical framework for adaptive photonic scale-up domains that makes this trade-off explicit and clarifies when reconfiguration is worthwhile. Along the way, we highlight a connection -- not surprising but still powerful -- between the Birkhoff--von Neumann (BvN) decomposition, maximum concurrent flow (a classic measure of network throughput), and the well-known $\alpha$-$\beta$ cost model for collectives. Finally, we outline a research agenda in algorithm design and systems integration that can build on this foundation.

cs.NI↗

arXiv2025

Short-circuiting Rings for Low-Latency AllReduce

Efficient collective communication is critical for many distributed ML and HPC applications. In this context, it is widely believed that the Ring algorithm for the AllReduce collective communication operation is optimal only for large messages, while Recursive Doubling is preferable for small ones due to its logarithmic number of steps compared to the linear number for Ring. In this paper, we challenge this long-held assumption and show that the Ring algorithm can remain optimal even for short messages in ring-based GPU-to-GPU topologies, once realistic propagation delays and link capacity constraints are accounted for. We find that the total propagation delay for both Ring and Recursive Doubling essentially sums to the same value, but the latter incurs significantly higher congestion due to longer hop counts, leading to increased completion times. This surprising result motivates our case for in-collective adaptive topologies, particularly in the context of emerging photonic interconnects, which can break through the limitations of static topology designs at the collective communication granularity. We design a \emph{simple and fast} heuristic for circuit-switching that enables Recursive Doubling to exploit dynamically reconfigurable photonic paths, carefully balancing reconfiguration delays, propagation latencies, and link congestion to minimize overall completion time. Our preliminary evaluations, using realistic reconfiguration delays, show that our circuit-switching schedules enable faster completion times for Recursive Doubling, even compared to Ring AllReduce on static ring topologies. We conclude by highlighting key challenges and future research directions for realizing practical, in-collective photonic switching.

cs.NI↗

arXiv2025

Vermilion: A Traffic-Aware Reconfigurable Optical Interconnect with Formal Throughput Guarantees

The increasing gap between datacenter traffic volume and the capacity of electrical switches has driven the development of reconfigurable network designs utilizing optical circuit switching. Recent advancements, particularly those featuring periodic fixed-duration reconfigurations, have achieved practical end-to-end delays of just a few microseconds. However, current designs rely on multi-hop routing to enhance utilization, which can lead to a significant reduction in worst-case throughput and added overhead from congestion control and routing complexity. These factors pose significant operational challenges for the large-scale deployment of these technologies. We present Vermilion, a reconfigurable optical interconnect that breaks the throughput barrier of existing periodic reconfigurable networks, without the need for multi-hop routing -- thus eliminating congestion control and simplifying routing to direct communication. Vermilion adopts a traffic-aware approach while retaining the simplicity of periodic fixed-duration reconfigurations, similar to RotorNet. We formally establish throughput bounds for Vermilion, demonstrating that it achieves at least $33\%$ more throughput in the worst-case compared to existing designs. The key innovation of Vermilion is its short traffic-aware periodic schedule, derived using a matrix rounding technique. This schedule is then combined with a traffic-oblivious periodic schedule to efficiently manage any residual traffic. Our evaluation results support our theoretical findings, revealing significant performance gains for datacenter workloads.

cs.NI↗

arXiv2024

Ethereal: Divide and Conquer Network Load Balancing in Large-Scale Distributed Training

Large-scale distributed training in production datacenters constitutes a challenging workload bottlenecked by network communication. In response, both major industry players (e.g., Ultra Ethernet Consortium) and parts of academia have surprisingly, and almost unanimously, agreed that packet spraying is \emph{necessary} to improve the performance of large-scale distributed training workloads. In this paper, we challenge this prevailing belief and pose the question: \emph{How close can singlepath transport come to matching the performance of packet spraying?} We demonstrate that singlepath transport (from a NIC's perspective) is sufficient and can perform nearly as well as ideal packet spraying, particularly in the context of distributed training in CLOS-based topologies. Our assertion is based on four key observations about workloads driven by collective communication patterns: \emph{(i)} flow sizes are known upon arrival, \emph{(ii)} flow sizes are equal within each step of a collective, \emph{(iii)} the completion time of a collective is more critical than individual flow completion times, and \emph{(iv)} flows can be \emph{split} upon arrival to control load balancing directly from the application layer. We present Ethereal, a simple distributed load balancing algorithm that opportunistically splits flows and assigns paths to each flow in a transparent manner, requiring little to no changes to existing RDMA NICs. Our evaluation, spanning a wide range of collective communication algorithms and GPT models using Astra-Sim, shows that Ethereal significantly reduces the completion times by up to $30\%$ compared to packet spraying and by up to $40\%$ compared to REPS, even under link failures. This paper offers an alternative perspective for developing next-generation transport protocols tailored to large-scale distributed training.

cs.NI↗

arXiv2024

Understanding the Throughput Bounds of Reconfigurable Datacenter Networks

The increasing gap between the growth of datacenter traffic volume and the capacity of electrical switches led to the emergence of reconfigurable datacenter network designs based on optical circuit switching. A multitude of research works, ranging from demand-oblivious (e.g., RotorNet, Sirius) to demand-aware (e.g., Helios, ProjecToR) reconfigurable networks, demonstrate significant performance benefits. Unfortunately, little is formally known about the achievable throughput of such networks. Only recently have the throughput bounds of demand-oblivious networks been studied. In this paper, we tackle a fundamental question: Whether and to what extent can demand-aware reconfigurable networks improve the throughput of datacenters? This paper attempts to understand the landscape of the throughput bounds of reconfigurable datacenter networks. Given the rise of machine learning workloads and collective communication in modern datacenters, we specifically focus on their typical communication patterns, namely uniform-residual demand matrices. We formally establish a separation bound of demand-aware networks over demand-oblivious networks, proving analytically that the former can provide at least $16\%$ higher throughput. Our analysis further uncovers new design opportunities based on periodic, fixed-duration reconfigurations that can harness the throughput benefits of demand-aware networks while inheriting the simplicity and low reconfiguration overheads of demand-oblivious networks. Finally, our evaluations corroborate the theoretical results of this paper, demonstrating that demand-aware networks significantly outperform oblivious networks in terms of throughput. This work barely scratches the surface and unveils several intriguing open questions, which we discuss at the end of this paper.

cs.NI↗

arXiv2024

Credence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions

Packet buffers in datacenter switches are shared across all the switch ports in order to improve the overall throughput. The trend of shrinking buffer sizes in datacenter switches makes buffer sharing extremely challenging and a critical performance issue. Literature suggests that push-out buffer sharing algorithms have significantly better performance guarantees compared to drop-tail algorithms. Unfortunately, switches are unable to benefit from these algorithms due to lack of support for push-out operations in hardware. Our key observation is that drop-tail buffers can emulate push-out buffers if the future packet arrivals are known ahead of time. This suggests that augmenting drop-tail algorithms with predictions about the future arrivals has the potential to significantly improve performance. This paper is the first research attempt in this direction. We propose Credence, a drop-tail buffer sharing algorithm augmented with machine-learned predictions. Credence can unlock the performance only attainable by push-out algorithms so far. Its performance hinges on the accuracy of predictions. Specifically, Credence achieves near-optimal performance of the best known push-out algorithm LQD (Longest Queue Drop) with perfect predictions, but gracefully degrades to the performance of the simplest drop-tail algorithm Complete Sharing when the prediction error gets arbitrarily worse. Our evaluations show that Credence improves throughput by $1.5$x compared to traditional approaches. In terms of flow completion times, we show that Credence improves upon the state-of-the-art approaches by up to $95\%$ using off-the-shelf machine learning techniques that are also practical in today's hardware. We believe this work opens several interesting future work opportunities both in systems and theory that we discuss at the end of this paper.

cs.NI↗

arXiv2022

Mars: Near-Optimal Throughput with Shallow Buffers in Reconfigurable Datacenter Networks

The performance of large-scale computing systems often critically depends on high-performance communication networks. Dynamically reconfigurable topologies, e.g., based on optical circuit switches, are emerging as an innovative new technology to deal with the explosive growth of datacenter traffic. Specifically, \emph{periodic} reconfigurable datacenter networks (RDCNs) such as RotorNet (SIGCOMM 2017), Opera (NSDI 2020) and Sirius (SIGCOMM 2020) have been shown to provide high throughput, by emulating a \emph{complete graph} through fast periodic circuit switch scheduling. However, to achieve such a high throughput, existing reconfigurable network designs pay a high price: in terms of potentially high delays, but also, as we show as a first contribution in this paper, in terms of the high buffer requirements. In particular, we show that under buffer constraints, emulating the high-throughput complete graph is infeasible at scale, and we uncover a spectrum of unvisited and attractive alternative RDCNs, which emulate regular graphs, but with lower node degree than the complete graph. We present Mars, a periodic reconfigurable topology which emulates a $d$-regular graph with near-optimal throughput. In particular, we systematically analyze how the degree~$d$ can be optimized for throughput given the available buffer and delay tolerance of the datacenter. We further show empirically that Mars achieves higher throughput compared to existing systems when buffer sizes are bounded.

cs.NI↗

arXiv2021

PowerTCP: Pushing the Performance Limits of Datacenter Networks

Increasingly stringent throughput and latency requirements in datacenter networks demand fast and accurate congestion control. We observe that the reaction time and accuracy of existing datacenter congestion control schemes are inherently limited. They either rely only on explicit feedback about the network state (e.g., queue lengths in DCTCP) or only on variations of state (e.g., RTT gradient in TIMELY). To overcome these limitations, we propose a novel congestion control algorithm, PowerTCP, which achieves much more fine-grained congestion control by adapting to the bandwidth-window product (henceforth called power). PowerTCP leverages in-band network telemetry to react to changes in the network instantaneously without loss of throughput and while keeping queues short. Due to its fast reaction time, our algorithm is particularly well-suited for dynamic network environments and bursty traffic patterns. We show analytically and empirically that PowerTCP can significantly outperform the state-of-the-art in both traditional datacenter topologies and emerging reconfigurable datacenters where frequent bandwidth changes make congestion control challenging. In traditional datacenter networks, PowerTCP reduces tail flow completion times of short flows by 80% compared to DCQCN and TIMELY, and by 33% compared to HPCC even at 60% network load. In reconfigurable datacenters, PowerTCP achieves 85% circuit utilization without incurring additional latency and cuts tail latency by at least 2x compared to existing approaches.

cs.NI↗

arXiv2021

Self-Adjusting Packet Classification

This paper is motivated by the vision of more efficient packet classification mechanisms that self-optimize in a demand-aware manner. At the heart of our approach lies a self-adjusting linear list data structure, where unlike in the classic data structure, there are dependencies, and some items must be in front of the others; for example, to correctly classify packets by rules arranged in a linked list, each rule must be in front of lower priority rules that overlap with it. After each access we can rearrange the list, similarly to Move-To-Front, but dependencies need to be respected. We present a 4-competitive online rearrangement algorithm, whose cost is at most four times worse than the optimal offline algorithm; no deterministic algorithm can be better than 3-competitive. The algorithm is simple and attractive, especially for memory-limited systems, as it does not require any additional memory (e.g., neither timestamps nor frequency statistics). Our approach can simply be deployed as a drop-in replacement for a static datastructure, potentially benefitting many existing networks. We evaluate our self-adjusting list packet classifier on realistic ruleset and traffic instances. We find that our classifier performs similarly to a static list for low-locality traffic, but significantly outperforms Efficuts (by a factor 7x), CutSplit (3.6x), and the static list (14x) for high locality and small rulesets. Memory consumption is 10x lower on average compared to Efficuts and CutSplit.

cs.DS↗

arXiv2021

FB: A Flexible Buffer Management Scheme for Data Center Switches

Today, network devices share buffer across priority queues to avoid drops during transient congestion. While cost-effective most of the time, this sharing can cause undesired interference among seemingly independent traffic. As a result, low-priority traffic can cause increased packet loss to high-priority traffic. Similarly, long flows can prevent the buffer from absorbing incoming bursts even if they do not share the same queue. The cause of this perhaps unintuitive outcome is that today's buffer sharing techniques are unable to guarantee isolation across (priority) queues without statically allocating buffer space. To address this issue, we designed FB, a novel buffer sharing scheme that offers strict isolation guarantees to high-priority traffic without sacrificing link utilizations. Thus, FB outperforms conventional buffer sharing algorithms in absorbing bursts while achieving on-par throughput. We show that FB is practical and runs at line-rate on existing hardware (Barefoot Tofino). Significantly, FB's operations can be approximated in non-programmable devices.

cs.NI↗

arXiv2021

Online List Access with Precedence Constraints

This paper considers a natural generalization of the online list access problem in the paid exchange model, where additionally there can be precedence constraints ("dependencies") among the nodes in the list. For example, this generalization is motivated by applications in the context of packet classification. Our main contributions are constant-competitive deterministic and randomized online algorithms, designed around a procedure Move-Recursively-Forward, a generalization of Move-To-Front tailored to handle node dependencies. Parts of the analysis build upon ideas of the classic online algorithms Move-To-Front and BIT, and address the challenges of the extended model. We further discuss the challenges related to insertions and deletions.

cs.DS↗

arXiv2020

Alias Resolution Based on ICMP Rate Limiting

Alias resolution techniques (e.g., Midar) associate, mostly through active measurement, a set of IP addresses as belonging to a common router. These techniques rely on distinct router features that can serve as a signature. Their applicability is affected by router support of the features and the robustness of the signature. This paper presents a new alias resolution tool called Limited Ltd. that exploits ICMP rate limiting, a feature that is increasingly supported by modern routers that has not previously been used for alias resolution. It sends ICMP probes toward target interfaces in order to trigger rate limiting, extracting features from the probe reply loss traces. It uses a machine learning classifier to designate pairs of interfaces as aliases. We describe the details of the algorithm used by Limited Ltd. and illustrate its feasibility and accuracy. Limited Ltd. not only is the first tool that can perform alias resolution on IPv6 routers that do not generate monotonically increasing fragmentation IDs (e.g., Juniper routers) but it also complements the state-of-the-art techniques for IPv4 alias resolution. All of our code and the collected dataset are publicly available.

cs.NI↗