arXiv subjects

Isaac Keslassy

Publications and source records attributed to Isaac Keslassy.

18 recordsLinked to original sources

arXiv2026

Incast-Free MoE Rate-Based Scheduling

Mixture of Experts (MoE) architectures have become key to large language models; however, their typical round-robin (RR) scheduling introduces significant bottlenecks. In this paper, we demonstrate that RR causes a previously-undiscovered exponential incast phenomenon with MoE traffic. We propose an alternative proactive fair scheduling framework tailored for MoE workloads, which effectively prevents fabric oversubscription. We also outline how it can be implemented in NICs. Finally, through extensive simulations with real and synthetic workloads, we demonstrate that this framework consistently eliminates incast, maintains a near-100% link utilization, and reduces Collective Completion Time (CCT).

cs.NI↗

arXiv2026

CAPS: Fine-Tuning CCA Timing

Data-center congestion control targets high throughput, fair bandwidth allocation, and low latency. Modern transports couple rate computation and packet scheduling into a single feedback loop, converging to near-optimal rates but leaving standing queues that can scale with the number of flows. We argue that separating the two reveals a simpler design point. Given stable feasible rates, the residual queue problem reduces to a timing problem: if every flow's packets arrive at the bottleneck in the correct slot, the link stays busy and the queue stays empty. Clocked ACK-Paced Synchronization CAPS is a lightweight distributed scheduling layer that achieves this by phase-locking each sender's transmissions to ACK-clocked bottleneck slots, with a per-flow correction that compensates for heterogeneous RTTs. We characterize the phase-locked steady state for dumbbell topologies under equal RTT, heterogeneous RTT, and bidirectional traffic, and validate the mechanism on a fat-tree under incast, permutation, and all-to-all traffic. CAPS reduces worst-case queue occupancy by 5-10x across all tested scenarios without throughput loss.

cs.NI↗

arXiv2026

Understanding Cross-Cloud Interconnects: Hands-On Measurements and Cost Optimization

New services such as Google Cross-Cloud Interconnect (CCI) address the rise in fast and large-scale cross-cloud data transfers. CCI offers dedicated high-throughput links with low per-GB transfer costs, but also involves high fixed leasing fees and multi-day provisioning delays. This combination makes cost optimization difficult because traffic patterns are unpredictable. This paper presents the first comprehensive study of CCI-like services. We begin with an empirical characterization of CCI and its alternatives using direct measurements across AWS-GCP interconnects. We then introduce ToggleCCI, a new dynamic cost-optimization algorithm designed to handle provisioning delays and uncertainty in future demand. ToggleCCI adapts by switching between VPN and CCI based on cost trends observed over a sliding time window. We prove that ToggleCCI achieves asymptotic optimality under sustained high-demand or low-demand regimes. Finally, using real-world traffic traces, we show that ToggleCCI consistently tracks the best static policy for each scenario and delivers substantial cost savings.

cs.NI↗

arXiv2026

Scheduling Parallel Optical Circuit Switches for AI Training

The rapid growth of AI training has dramatically increased datacenter traffic demand and energy consumption, which has motivated renewed interest in optical circuit switches (OCSes) as a high-bandwidth, energy-efficient alternative for AI fabrics. Deploying multiple parallel OCSes is a leading alternative. However, efficiently scheduling time-varying traffic matrices across parallel optical switches with non-negligible reconfiguration delays remains an open challenge. We consider the problem of scheduling a single AI traffic demand matrix $D$ over $s$ parallel OCSes while minimizing the makespan under reconfiguration delay $\delta$. Our algorithm Spectra relies on a three-step approach: Decompose $D$ into a minimal set of weighted permutations; Schedule these permutations across parallel switches using load-aware assignment; then Equalize the imbalanced loads on the switches via controlled permutation splitting. Evaluated on realistic AI training workloads (GPT model and Qwen MoE expert routing) as well as standard benchmarks, Spectra vastly outperforms a baseline based on state-of-the-art algorithms, reducing schedule makespan by an average factor of $1.4\times$ on GPT AI workloads, $1.9\times$ on MoE AI workloads, and $2.4\times$ on standard benchmarks. Further, the makespans achieved by Spectra consistently approach newly derived lower bounds.

cs.NI↗

arXiv2026

CrossCheck: Input Validation for WAN Control Systems

We present CrossCheck, a system that validates inputs to the Software-Defined Networking (SDN) controller in a Wide Area Network (WAN). By detecting incorrect inputs - often stemming from bugs in the SDN control infrastructure - CrossCheck alerts operators before they trigger network outages. Our analysis at a large-scale WAN operator identifies invalid inputs as a leading cause of major outages, and we show how CrossCheck would have prevented those incidents. We deployed CrossCheck as a shadow validation system for four weeks in a production WAN, during which it accurately detected the single incident of invalid inputs that occurred while sustaining a 0% false positive rate under normal operation, hence imposing little additional burden on operators. In addition, we show through simulation that CrossCheck reliably detects a wide range of invalid inputs (e.g., detecting demand perturbations as small as 5% with 100% accuracy) and maintains a near-zero false positive rate for realistic levels of noisy, missing, or buggy telemetry data (e.g., sustaining zero false positives with up to 30% of corrupted telemetry data).

cs.NI↗

arXiv2026

Scaling Routers with In-Package Optics and High-Bandwidth Memories

This paper aims to apply two major scaling transformations from the computing packaging industry to internet routers: the heterogeneous integration of high-bandwidth memories (HBMs) and chiplets, as well as in-package optics. We propose a novel internet router architecture that employs these technologies to achieve a petabit/sec router within a single integrated package. At the top-level, we introduce a novel split-parallel switch architecture that spatially divides (without processing) the incoming fibers and distributes them across smaller independent switches without intermediate OEO conversions or fine-tuned per-packet load-balancing. This passive spatial division enables scaling at the cost of a coarser traffic load balancing. Yet, through extensive evaluations of backbone network traffic, we demonstrate that differences with fine-tuned approaches are small. In addition, we propose a novel HBM-based shared-memory architecture for the implementation of the smaller independent switches, and we introduce a novel parallel frame interleaving algorithm that packs traffic into frames so that HBM banks are accessed at peak HBM data rates in a cyclical interleaving manner. We further discuss why these new technologies represent a paradigm shift in the design of future internet routers. Finally, we emphasize that power consumption may constitute the primary bottleneck to scaling.

cs.NI↗

arXiv2025

Congestion Control for Spraying with Congested Paths

Packet spraying approaches are increasingly deployed in datacenter networks. However, their combination with existing congestion control algorithms (CCAs) may lead to poor QoS, especially when some of the paths are congested. In this paper, we first model the throughput collapse of a wide array of CCAs when some of the paths are congested. We explain that since CCAs are typically designed for single-path routing, their estimation function focuses on the latest feedback and mishandles feedback that reflects multiple paths. We propose using a median feedback that is more robust to the varying signals that come with multiple paths. We introduce MSwift and MNSCC, which apply this median principle to Google's Swift and Ultra Ethernet's NSCC. We demonstrate that they can improve both CCAs, reaching better QoS both under congested paths and in uncongested networks.

cs.NI↗

arXiv2025

2SYN: Congestion-Aware Multihoming

When sending flows to arbitrary destinations, current multihoming routers adopt simple congestion-oblivious mechanisms. Therefore, they cannot avoid congested paths. In this paper, we introduce 2SYN, the first congestion-aware multihoming algorithm that works for any destination. We explain how it dynamically selects a preferred path for new connections, even given previously-unseen destinations. We further demonstrate that it can be easily implemented in Linux. Finally, in a real-world experiment with either LTE or a wired link, we show how 2SYN dynamically adapts to the quality of the connection and outperforms alternative approaches. Thus, 2SYN helps companies better manage their networks by leveraging their multihoming capabilities.

cs.NI↗

arXiv2025

Load Balancing for AI Training Workloads

The extreme bandwidth demands of AI training has made load-balancing a critical component in AI fabrics, and a variety of load-balancing designs have emerged in recent work from both industry and research. However, there is currently little consensus on which design approach dominates or the conditions under which an approach dominates. We also lack an understanding of how far these approaches are from optimal. We provide a technical foundation for answering these questions by systematically evaluating leading load-balancing designs, while decoupling them from specific congestion control and loss recovery stacks. We find that load-balancing based on packet spraying dominates traditional approaches that load balance traffic at flow, flowlet, or subflow granularities. When comparing host- vs switch-based approaches to packet spraying, we find that they perform similarly in failure-free scenarios but that a host-based approach dominates under link failure because of its rapid visibility into end-to-end path conditions. We also identify that no leading approach achieves optimal O(1) queue scaling at maximum utilization. We demonstrate why a destination-based rotation (DR) discipline can reach this optimum and introduce Ofan, a switch-based implementation of DR that we show offers valuable performance gains over other packet spraying approaches.

cs.NI↗

arXiv2021

On the Persistent-Idle Load Distribution Policy Under Batch Arrivals and Random Service Capacity

The Persistent-Idle (PI) load distribution policy was recently introduced as an appealing alternative to current low-communication load balancing techniques. In PI, servers only update the dispatcher when they become idle, and the dispatcher always sends jobs to the last server that reported being idle. PI is unique in that it does not seek to push the server queue lengths towards equalization greedily. Rather, it aggressively pulls the servers away from starvation. As a result, PI's analysis requires different tools than other load balancing approaches. So far, PI was proven to achieve the stability region for Bernoulli arrivals and deterministic and constant service capacities. Our main contribution is proving that PI achieves the stability region in a model with batch arrivals and random service capacities. Proving this result requires developing tighter bounds on quantities of interest and proving the fulfillment of a generalized version of the state-dependent drift criteria previously considered. We also present PI-Split, which allows batches to be split among idle servers and prove its stability. Finally, we conduct simulations comparing PI and PI-Split to other load balancing policies. Simulation results indicate that both achieve excellent performance for the model under consideration for a wide range of parameters.

math.PR↗

arXiv2020

AnchorHash: A Scalable Consistent Hash

Consistent hashing (CH) is a central building block in many networking applications, from datacenter load-balancing to distributed storage. Unfortunately, state-of-the-art CH solutions cannot ensure full consistency under arbitrary changes and/or cannot scale while maintaining reasonable memory footprints and update times. We present AnchorHash, a scalable and fully-consistent hashing algorithm. AnchorHash achieves high key lookup rates, a low memory footprint, and low update times. We formally establish its strong theoretical guarantees, and present advanced implementations with a memory footprint of only a few bytes per resource. Moreover, extensive evaluations indicate that it outperforms state-of-the-art algorithms, and that it can scale on a single core to 100 million resources while still achieving a key lookup rate of more than 15 million keys per second.

cs.DS↗

arXiv2020

LSQ: Load Balancing in Large-Scale Heterogeneous Systems with Multiple Dispatchers

Nowadays, the efficiency and even the feasibility of traditional load-balancing policies are challenged by the rapid growth of cloud infrastructure and the increasing levels of server heterogeneity. In such heterogeneous systems with many load-balancers, traditional solutions, such as JSQ, incur a prohibitively large communication overhead and detrimental incast effects due to herd behavior. Alternative low-communication policies, such as JSQ(d) and the recently proposed JIQ, are either unstable or provide poor performance. We introduce the Local Shortest Queue (LSQ) family of load balancing algorithms. In these algorithms, each dispatcher maintains its own, local, and possibly outdated view of the server queue lengths, and keeps using JSQ on its local view. A small communication overhead is used infrequently to update this local view. We formally prove that as long as the error in these local estimates of the server queue lengths is bounded in expectation, the entire system is strongly stable. Finally, in simulations, we show how simple and stable LSQ policies exhibit appealing performance and significantly outperform existing low-communication policies, while using an equivalent communication budget. In particular, our simple policies often outperform even JSQ due to their reduction of herd behavior. We further show how, by relying on smart servers (i.e., advanced pull-based communication), we can further improve performance and lower communication overhead.

cs.NI↗

arXiv2020

RADE: Resource-Efficient Supervised Anomaly Detection Using Decision Tree-Based Ensemble Methods

Decision-tree-based ensemble classification methods (DTEMs) are a prevalent tool for supervised anomaly detection. However, due to the continued growth of datasets, DTEMs result in increasing drawbacks such as growing memory footprints, longer training times, and slower classification latencies at lower throughput. In this paper, we present, design, and evaluate RADE - a DTEM-based anomaly detection framework that augments standard DTEM classifiers and alleviates these drawbacks by relying on two observations: (1) we find that a small (coarse-grained) DTEM model is sufficient to classify the majority of the classification queries correctly, such that a classification is valid only if its corresponding confidence level is greater than or equal to a predetermined classification confidence threshold; (2) we find that in these fewer harder cases where our coarse-grained DTEM model results in insufficient confidence in its classification, we can improve it by forwarding the classification query to one of expert DTEM (fine-grained) models, which is explicitly trained for that particular case. We implement RADE in Python based on scikit-learn and evaluate it over different DTEM methods: RF, XGBoost, AdaBoost, GBDT and LightGBM, and over three publicly available datasets. Our evaluation over both a strong AWS EC2 instance and a Raspberry Pi 3 device indicates that RADE offers competitive and often superior anomaly detection capabilities as compared to standard DTEM methods, while significantly improving memory footprint (by up to 5.46x), training-time (by up to 17.2x), and classification latency (by up to 31.2x).

cs.LG↗

arXiv2018

Pied Piper: Rethinking Internet Data Delivery

We contend that, analogously to the transition from resource-limited on-prem computing to resource-abundant cloud computing, Internet data delivery should also be adapted to a reality in which the cloud offers a virtually unlimited resource, i.e., network capacity, and virtualization enables delegating local tasks, such as routing and congestion control, to the cloud. This necessitates rethinking the traditional roles of inter- and intra-domain routing and conventional end-to-end congestion control. We introduce Optimized Cloudified Delivery (OCD), a holistic approach for optimizing joint Internet/cloud data delivery, and evaluate OCD through hundreds of thousands of file downloads from multiple locations. We start by examining an OCD baseline approach: traffic from a source A to a destination B successively passes through two cloud virtual machines operating as relays - nearest to A and B; and the two cloud relays employ TCP split. We show that even this naive strategy can outperform recently proposed improved end-to-end congestion control paradigms (BBR and PCC) by an order of magnitude. Next, we present a protocol-free, ideal pipe model of data transmission, and identify where today's Internet data delivery mechanisms diverge from this model. We then design and implement OCD Pied Piper. Pied Piper leverages various techniques, including novel kernel-based transport-layer accelerations, to improve the Internet-Cloud interface so as to approximately match the ideal network pipe model.

cs.NI↗

arXiv2018

Memento: Making Sliding Windows Efficient for Heavy Hitters

Cloud operators require real-time identification of Heavy Hitters (HH) and Hierarchical Heavy Hitters (HHH) for applications such as load balancing, traffic engineering, and attack mitigation. However, existing techniques are slow in detecting new heavy hitters. In this paper, we make the case for identifying heavy hitters through \textit{sliding windows}. Sliding windows detect heavy hitters quicker and more accurately than current methods, but to date had no practical algorithms. Accordingly, we introduce, design and analyze the \textit{Memento} family of sliding window algorithms for the HH and HHH problems in the single-device and network-wide settings. Using extensive evaluations, we show that our single-device solutions attain similar accuracy and are by up to $273\times$ faster than existing window-based techniques. Furthermore, we exemplify our network-wide HHH detection capabilities on a realistic testbed. To that end, we implemented Memento as an open-source extension to the popular HAProxy cloud load-balancer. In our evaluations, using an HTTP flood by 50 subnets, our network-wide approach detected the new subnets faster, and reduced the number of undetected flood requests by up to $37\times$ compared to the alternatives.

cs.NI↗

arXiv2016

Links as a Service (LaaS): Feeling Alone in the Shared Cloud

The most demanding tenants of shared clouds require complete isolation from their neighbors, in order to guarantee that their application performance is not affected by other tenants. Unfortunately, while shared clouds can offer an option whereby tenants obtain dedicated servers, they do not offer any network provisioning service, which would shield these tenants from network interference. In this paper, we introduce Links as a Service, a new abstraction for cloud service that provides physical isolation of network links. Each tenant gets an exclusive set of links forming a virtual fat tree, and is guaranteed to receive the exact same bandwidth and delay as if it were alone in the shared cloud. Under simple assumptions, we derive theoretical conditions for enabling LaaS without capacity over-provisioning in fat-trees. New tenants are only admitted in the network when they can be allocated hosts and links that maintain these conditions. Using experiments on real clusters as well as simulations with real-life tenant sizes, we show that LaaS completely avoids the performance degradation caused by traffic from concurrent tenants on shared links. Compared to mere host isolation, LaaS can improve the application performance by up to 200%, at the cost of a 10% reduction in the cloud utilization.

cs.DC↗

arXiv2011

Maximum Bipartite Matching Size And Application to Cuckoo Hashing

Cuckoo hashing with a stash is a robust multiple choice hashing scheme with high memory utilization that can be used in many network device applications. Unfortunately, for memory loads beyond 0.5, little is known on its performance. In this paper, we analyze its average performance over such loads. We tackle this problem by recasting the problem as an analysis of the expected maximum matching size of a given random bipartite graph. We provide exact results for any finite system, and also deduce asymptotic results as the memory size increases. We further consider other variants of this problem, and finally evaluate the performance of our models on Internet backbone traces. More generally, our results give a tight lower bound on the size of the stash needed for any multiple-choice hashing scheme.

cs.DS↗

arXiv2011

Multi-Amdahl: Optimal Resource Sharing with Multiple Program Execution Segments

This paper presents Multi-Amdahl, a resource allocation analytical tool for heterogeneous systems. Our model includes multiple program execution segments, where each one is accelerated by a specific hardware unit. The acceleration speedup of the specific hardware unit is a function of a limited resource, such as the unit area, power, or energy. Using the Lagrange theorem we discover the optimal resource distribution between all specific units. We then illustrate this general Multi-Amdahl technique using several examples of area and power allocation among several cores and accelerators.

cs.AR↗