arXiv subjects

Albert Gran Alcoz

Publications and source records attributed to Albert Gran Alcoz.

3 recordsLinked to original sources

arXiv2026

High-speed Networking for Giga-Scale AI Factories

As distributed model training scales to span hundreds of thousands of GPUs, scale-out networks face unprecedented performance and efficiency demands. NVIDIA Spectrum-X Ethernet has been designed from the ground up to achieve predictable and stable network performance with high utilization and low latency. This paper presents the Spectrum-X multiplane architecture, which replaces hierarchical depth with topological parallelism, and introduces hardware-accelerated load balancing in NICs and switches as the key architectural approach to provide fast reaction to highly dynamic network conditions at the microsecond timescales that AI training workloads demand. We describe the motivation, design principles, evaluation methodology and performance on state-of-the-art benchmarks, as well as the lessons we learned from deploying and debugging Spectrum-X networks in large-scale systems. Our evaluation highlights production-grade AI infrastructure performance across three core dimensions: 98% of the theoretical line rate with low jitter-free latency; strong cross-tenant isolation for concurrent workloads; robust, capacity-proportional bisection bandwidth and 7% latency increase for 10% fabric link failures; and rapid reaction to host and fabric link flaps during LLM training workloads.

cs.NI↗

arXiv2024

FitNets: An Adaptive Framework to Learn Accurate Traffic Distributions

Learning precise distributions of traffic features (e.g., burst sizes, packet inter-arrival time) is still a largely unsolved problem despite being critical for management tasks such as capacity planning or anomaly detection. A key limitation nowadays is the lack of feedback between the control plane and the data plane. Programmable data planes offer the opportunity to create systems that let data- and control plane to work together, compensating their respective shortcomings. We present FitNets, an adaptive network monitoring system leveraging feedback between the data- and the control plane to learn accurate traffic distributions. In the control plane, FitNets relies on Kernel Density Estimators which allow to provably learn distributions of any shape. In the data plane, FitNets tests the accuracy of the learned distributions while dynamically adapting data collection to the observed distribution fitness, prioritizing under-fitted features. We have implemented FitNets in Python and P4 (including on commercially available programmable switches) and tested it on real and synthetic traffic traces. FitNets is practical: it is able to estimate hundreds of distributions from up to 60 millions samples per second, while providing accurate error estimates and adapting to complex traffic patterns.

cs.NI↗

arXiv2023

Everything Matters in Programmable Packet Scheduling

Programmable packet scheduling allows the deployment of scheduling algorithms into existing switches without need for hardware redesign. Scheduling algorithms are programmed by tagging packets with ranks, indicating their desired priority. Programmable schedulers then execute these algorithms by serving packets in the order described in their ranks. The ideal programmable scheduler is a Push-In First-Out (PIFO) queue, which achieves perfect packet sorting by pushing packets into arbitrary positions in the queue, while only draining packets from the head. Unfortunately, implementing PIFO queues in hardware is challenging due to the need to arbitrarily sort packets at line rate based on their ranks. In the last years, various techniques have been proposed, approximating PIFO behaviors using the available resources of existing data planes. While promising, approaches to date only approximate one of the characteristic behaviors of PIFO queues (i.e., its scheduling behavior, or its admission control). We propose PACKS, the first programmable scheduler that fully approximates PIFO queues on all their behaviors. PACKS does so by smartly using a set of strict-priority queues. It uses packet-rank information and queue-occupancy levels at enqueue to decide: whether to admit packets to the scheduler, and how to map admitted packets to the different queues. We fully implement PACKS in P4 and evaluate it on real workloads. We show that PACKS: better-approximates PIFO than state-of-the-art approaches and scales. We also show that PACKS runs at line rate on existing hardware (Intel Tofino).

cs.NI↗