arXiv subjects

Daniel Amir

Publications and source records attributed to Daniel Amir.

5 recordsLinked to original sources

arXiv2026

MoX: Efficient MoE Routing on Direct-Connect Topologies

Optically switched networks suit the regular communication of dense ML models, but MoE introduces sparse, runtime-dependent traffic. We show that efficient offline-optimized routing enables efficient MoE training and inference on direct-connect topologies without the need for MoE traffic matrix or dynamic topology reconfiguration. MoX constructs token-aware multicast trees to reduce bandwidth tax, then uses static, precomputed link weights to balance traffic by solving a restricted multicast tree-packing problem. Using recorded traffic from large MoE models, token-level traces, and ASTRA-sim, we find that MoX accelerates the full MoE block -- dispatch, expert computation, and combine -- by up to 1.8x over min-hop routing. Moreover, it attains nearly ideal packet-switched network performance in random expander topologies. On a 1,024-TPU model of Google's Boardfly topology, MoX reduces the dispatch bottleneck link load by up to 47%. These results show that high-performance MoE on static direct-connect fabrics can be achieved via optimized load-oblivious routing without demand-driven reconfiguration.

cs.NI↗

arXiv2026

SprayCheck: Finding Gray Failures in Adaptive Routing Networks

Distributed machine learning (ML) training has become a dominant workload in modern data center networks, operating at massive scale with clusters comprising tens to hundreds of thousands of GPUs. The scale of these networks makes failures, and particularly gray failures, inevitable. Gray failures can significantly degrade both network and application performance, yet they are notoriously difficult to detect, localize, and debug. To meet the performance demands of ML workloads, adaptive routing is widely deployed to maximize network utilization by dynamically spreading traffic across many paths. While adaptive routing increases network utilization, it also greatly intensifies the effect of gray failures. Prior work has either dismissed gray failures as negligible or proposed detection mechanisms that fail to scale, rendering these approaches increasingly impractical for large-scale clusters. We present SprayCheck, a passive gray failure detection system that leverages the statistical properties of adaptive routing and network load balancing. By combining these properties with flow-level information, SprayCheck can identify failures before they significantly impact application performance, enabling preemptive rerouting and improving overall performance. Importantly, this is achieved through passive observation of traffic spraying, without introducing additional load on the network. We evaluate SprayCheck and show that it can detect and localize a single-link packet-drop-rate $1.5\%$ within a single iteration and as little as $0.5\%$ within 5 training iterations of Llama-3 70B in a 64 spine topology.

cs.NI↗

arXiv2026

ACOS: Arrays of Cheap Optical Switches

Machine learning training places immense demands on cluster networks, motivating specialized architectures and co-design with parallelization strategies. Recent designs incorporating optical circuit switches (OCSes) are promising, offering improved cost, power efficiency, and long-term bandwidth scaling than packet switches. However, most existing approaches rely on costly high-radix OCSes and/or combine them with packet switches to achieve competitive performance at scale. Unfortunately, high-radix OCSes are both expensive and slow to reconfigure, limiting both scalability and performance. We propose Arrays of Cheap Optical Switches (ACOS), which bring application co-design directly to the structure of the reconfigurable fabric. Using low-radix OCSes as building blocks, ACOS supports the forms of reconfiguration needed in training clusters including topology selection, workload adaptation, and failure resilience. The cost of ACOS scales with supported topologies and adaptations rather than with port count, breaking past the scalability barriers of current specialized ML networks. We show through simulation that ACOS-based deployments match the performance of fully provisioned packet-switched networks when training state-of-the-art LLMs at scale, while delivering significant cost savings using existing off-the-shelf OCSes, with strong bandwidth scaling and higher cost savings in the future.

cs.NI↗

arXiv2023

Breaking the VLB Barrier for Oblivious Reconfigurable Networks

In a landmark 1981 paper, Valiant and Brebner gave birth to the study of oblivious routing and, simultaneously, introduced its most powerful and ubiquitous method: Valiant load balancing (VLB). By routing messages through a randomly sampled intermediate node, VLB lengthens routing paths by a factor of two but gains the crucial property of obliviousness. Forty years later, with datacenters handling workloads whose communication pattern varies too rapidly to allow centralized coordination, VLB continues to take center stage as a widely used - and in some cases provably optimal - way to balance load in the network obliviously to the traffic demands. However, the ability of the network to rapidly reconfigure its interconnection topology gives rise to new possibilities. In this work we revisit the question of whether VLB remains optimal in the novel setting of reconfigurable networks. Prior work showed that VLB achieves the optimal tradeoff between latency and guaranteed throughput. In this work we show that a strictly superior latency-throughput tradeoff is achievable when the throughput bound is relaxed to hold with high probability. The same improved tradeoff is also achievable with guaranteed throughput under time-stationary demands, provided the latency bound is relaxed to hold with high probability and that the network is allowed to be semi-oblivious, using an oblivious (randomized) connection schedule but demand-aware routing. We prove that the latter result is not achievable by any fully-oblivious reconfigurable network design, marking a rare case in which semi-oblivious routing has a provable asymptotic advantage over oblivious routing. To analyze our routing scheme we prove an exponential tail bound which may be of independent interest, concerning the distribution of values of a bilinear form on an orbit of a permutation group action.

cs.DS↗

arXiv2021

Optimal Oblivious Reconfigurable Networks

Oblivious routing has a long history in both the theory and practice of networking. In this work we initiate the formal study of oblivious routing in the context of reconfigurable networks, a new architecture that has recently come to the fore in datacenter networking. These networks allow a rapidly changing bounded-degree pattern of interconnections between nodes, but the network topology and the selection of routing paths must both be oblivious to the traffic demand matrix. Our focus is on the trade-off between maximizing throughput and minimizing latency in these networks. For every constant throughput rate, we characterize (up to a constant factor) the minimum latency achievable by an oblivious reconfigurable network design that satisfies the given throughput guarantee. The trade-off between these two objectives turns out to be surprisingly subtle: the curve depicting it has an unexpected scalloped shape reflecting the fact that load-balancing becomes more difficult when the average length of routing paths is not an integer because equalizing all the path lengths is not possible. The proof of our lower bound uses LP duality to verify that Valiant load balancing is the most efficient oblivious routing scheme when used in combination with an optimally-designed reconfigurable network topology. The proof of our upper bound uses an algebraic construction in which the network nodes are identified with vectors over a finite field, the network topology is described by either the elementary basis or a sequence of Vandermonde matrices, and routing paths are constructed by selecting columns of these matrices to yield the appropriate mixture of path lengths within the shortest possible time interval.

cs.DS↗