Next-generation datacenters require highly efficient network load balancing to manage the growing scale of artificial intelligence (AI) training and general datacenter traffic. However, existing Ethernet-based solutions, such as Equal Cost Multi-Path (ECMP) and oblivious packet spraying (OPS), struggle to maintain high network utilization due to both increasing traffic demands and the expanding scale of datacenter topologies, which also exacerbate network failures. To address these limitations, we propose REPS, a lightweight decentralized per-packet adaptive load balancing algorithm designed to optimize network utilization while ensuring rapid recovery from link failures. REPS adapts to network conditions by caching good-performing paths. In case of a network failure, REPS re-routes traffic away from it in less than 100 microseconds. REPS is designed to be deployed with next-generation out-of-order transports, such as Ultra Ethernet, and uses less than 25 bytes of per-connection state regardless of the topology size. We extensively evaluate REPS in large-scale simulations and FPGA-based NICs.

Network latency severely impacts the performance of applications running on supercomputers. Adaptive routing algorithms route packets over different available paths to reduce latency and improve network utilization. However, if a switch routes packets belonging to the same network flow on different paths, they might arrive at the destination out-of-order due to differences in the latency of these paths. For some transport protocols like TCP, QUIC, and RoCE, out-of-order (OOO) packets might cause large performance drops or significantly increase CPU utilization. In this work, we proposeFlowcut switching, a new adaptive routing algorithm that provides high-performance in-order packet delivery. Differently from existing solutions likeFlowlet switching, which are based on the assumption of bursty traffic and that might still reorder packets,Flowcut switchingguarantees in-order delivery under any network conditions, and is effective also for non-bursty traffic, as it is often the case for RDMA. On top of this, Flowcut can be implemented either at the switch or NIC level providing flexibility and different tradeoffs.

NET4EXA aims to develop a next-generation high-performance interconnect for HPC and AI systems, addressing the increasing demands of large-scale infrastructures, such as those required for training Large Language Models. Building upon the proven BXI (Bull eXascale Interconnect) European technology used in TOP15 supercomputers, NET4EXA will deliver the new BXI release, BXIv3, a complete hardware and software interconnect solution, including switch and network interface components. The project will integrate a fully functional pilot system at TRL 8, ready for deployment into upcoming exascale and post-exascale systems from 2025 onward. Leveraging prior research from European initiatives like RED-SEA, the previous achievements of consortium partners and over 20 years of expertise from BULL, NET4EXA also lays the groundwork for the future generation of BXI, BXIv4, providing analysis and preliminary design. The project will use a hybrid development and co-design approach, combining commercial switch technology with custom IP and FPGA-based NICs. Performances of NET4EXA BXIv3 interconnect will be evaluated using a broad portfolio of benchmarks, scientific scalable applications, and AI workloads.

Cloud computing and AI workloads are driving unprecedented demand for efficient communication within and across datacenters. However, the coexistence of intra- and inter-datacenter traffic within datacenters plus the disparity between the RTTs of intra- and inter-datacenter networks complicates congestion management and traffic routing. Particularly, faster congestion responses of intra-datacenter traffic causes rate unfairness when competing with slower inter-datacenter flows. Additionally, inter-datacenter messages suffer from slow loss recovery and, thus, require reliability. Existing solutions overlook these challenges and handle inter- and intra-datacenter congestion with separate control loops or at different granularities. We propose Uno, a unified system for both inter- and intra-DC environments that integrates a transport protocol for rapid congestion reaction and fair rate control with a load balancing scheme that combines erasure coding and adaptive routing. Our findings show that Uno significantly improves the completion times of both inter- and intra-DC flows compared to state-of-the-art methods such as Gemini.

Communication locality plays a key role in the performance of collective operations on large HPC systems, especially on oversubscribed networks where groups of nodes are fully connected internally but sparsely linked through global connections. We present Bine (binomial negabinary) trees, a family of collective algorithms that improve communication locality. Bine trees maintain the generality of binomial trees and butterflies while cutting global-link traffic by up to \(33\%\). We implement eight Bine-based collectives and evaluate them on four large-scale supercomputers with Dragonfly, Dragonfly+, oversubscribed fat-tree, and torus topologies, achieving up to 5 × speedups and consistent reductions in global-link traffic across different vector sizes and node counts.

Efficient Reduce and AllReduce communication collectives are a critical cornerstone of high-performance computing (HPC) applications. We present the first systematic investigation of Reduce and AllReduce on the Cerebras Wafer-Scale Engine (WSE). This architecture has been shown to achieve unprecedented performance both for machine learning workloads and other computational problems like FFT. We introduce a performance model to estimate the execution time of algorithms on the WSE and validate our predictions experimentally for a wide range of input sizes. In addition to existing implementations, we design and implement several new algorithms specifically tailored to the architecture. Moreover, we establish a lower bound for the runtime of a Reduce operation on the WSE. Based on our model, we automatically generate code that achieves near-optimal performance across the whole range of input sizes. Experiments demonstrate that our new Reduce and AllReduce algorithms outperform the current vendor solution by up to 3.27×. Additionally, our model predicts performance with less than 4% error. The proposed communication collectives increase the range of HPC applications that can benefit from the high throughput of the WSE. Our model-driven methodology demonstrates a disciplined approach that can lead the way to further algorithmic advancements on wafer-scale architectures.

Multi-GPU nodes are increasingly common in the rapidly evolving landscape of exascale supercomputers. On these systems, GPUs on the same node are connected through dedicated networks, with bandwidths up to a few terabits per second. However, gauging performance expectations and maximizing system efficiency is challenging due to different technologies, design options, and software layers. This paper comprehensively characterizes three supercomputers — Alps, Leonardo, and LUMI — each with a unique architecture and design. We focus on performance evaluation of intra-node and inter-node interconnects on up to 4,096 GPUs, using a mix of intra-node and inter-node benchmarks. By analyzing its limitations and opportunities, we aim to offer practical guidance to researchers, system architects, and software developers dealing with multi-GPU supercomputing. Our results show that there is untapped bandwidth, and there are still many opportunities for optimization, ranging from network to software optimization.

The allreduce operation is an essential building block for many distributed applications, ranging from the training of deep learning models to scientific computing. In an allreduce operation, data from multiple hosts is aggregated together and then broadcasted to each host participating in the operation. Allreduce performance can be improved by a factor of two by aggregating the data directly in the network. Switches aggregate data coming from multiple ports before forwarding the partially aggregated result to the next hop. In all existing solutions, each switch needs to know the ports from which it will receive the data to aggregate. However, this forces packets to traverse a predefined set of switches, making these solutions prone to congestion. For this reason, we design Canary, the first congestion-aware in-network allreduce algorithm. Canary uses load balancing algorithms to forward packets on the least congested paths. Because switches do not know from which ports they will receive the data to aggregate, they use timeouts to aggregate the data in a best-effort way. We develop a P4 Canary prototype and evaluate it on a Tofino switch. We then validate Canary through simulations on large networks, showing performance improvements up to 40% compared to the state-of-the-art.

This paper presents a security analysis of the InfiniBand architecture, a prevalent RDMA standard, and NVMe-over-Fabrics (NVMe-oF), a prominent protocol for industrial disaggregated storage that exploits RDMA protocols to achieve low-latency and high-bandwidth access to remote solid-state devices. Our work, NeVerMore, discovers new vulnerabilities in RDMA protocols that unveils several attack vectors on RDMA-enabled applications and the NVMe-oF protocol, showing that the current security mechanisms of the NVMe-oF protocol do not address the security vulnerabilities posed by the use of RDMA. In particular, we show how an unprivileged user can inject packets into any RDMA connection created on a local network controller, bypassing security mechanisms of the operating system and its kernel, and how the injection can be used to acquire unauthorized block access to NVMe-oF devices. Overall, we implement four attacks on RDMA protocols and seven attacks on the NVMe-oF protocol and verify them on the two most popular implementations of NVMe-oF: SPDK and the Linux kernel. To mitigate the discovered attacks we propose multiple mechanisms that can be implemented by RDMA and NVMe-oF providers.

High-performance clusters and datacenters pose increasingly demanding requirements on storage systems. If these systems do not operate at scale, applications are doomed to become I/O bound and waste compute cycles. To accelerate the data path to remote storage nodes, remote direct memory access (RDMA) has been embraced by storage systems to let data flow from the network to storage targets, reducing overall latency and CPU utilization. Yet, this approach still involves CPUs on the data path to enforce storage policies such as authentication, replication, and erasure coding. We show how storage policies can be offloaded to fully programmable SmartNICs, without involving host CPUs. By using PsPIN, an open-hardware SmartNIC, we show latency improvements for writes (up to 2x), data replication (up to 2x), and erasure coding (up to 2x), when compared to respective CPU- and RDMA-based alternatives.

Numerous microarchitectural optimizations unlocked tremendous processing power for deep neural networks that in turn fueled the AI revolution. With the exhaustion of such optimizations, the growth of modern AI is now gated by the performance of training systems, especially their data movement. Instead of focusing on single accelerators, we investigate data-movement characteristics of large-scale training at full system scale. Based on our workload analysis, we design HammingMesh, a novel network topology that provides high bandwidth at low cost with high job scheduling flexibility. Specifically, HammingMesh can support full bandwidth and isolation to deep learning training jobs with two dimensions of parallelism. Furthermore, it also supports high global bandwidth for generic traffic. Thus, HammingMesh will power future large-scale deep learning systems with extreme bandwidth requirements.

In fault tolerance for parallel and distributed systems, message logging protocols have played a prominent role in the last three decades. Such protocols enable local rollback to provide recovery from fail-stop errors. Global rollback techniques can be straightforward to implement but at times lead to slower recovery than local rollback. Local rollback is more complicated but can offer faster recovery times. In this work, we study the power and energy efficiency implications of global and local rollback. We propose a power-efficient version of local rollback to reduce power consumption for non-critical,blockedprocesses, usingDynamic Voltage and Frequency Scaling(DVFS) andclock modulation(CM). Our results for 3 different MPI codes on 2 parallel systems show that power-efficient local rollback reduces CPU energy waste up to 50% during the recovery phase, compared to existing global and local rollback techniques, without introducing significant overheads. Furthermore, we show that savings manifest for all blocked processes, which grow linearly with the process count. We estimate that for settings with high recovery overheads the total energy waste of parallel codes is reduced with the proposed local rollback.

The allreduce operation is one of the most commonly used communication routines in distributed applications. To improve its bandwidth and to reduce network traffic, this operation can be accelerated by offloading it to network switches, that aggregate the data received from the hosts, and send them back the aggregated result. However, existing solutions provide limited customization opportunities and might provide suboptimal performance when dealing with custom operators and data types, with sparse data, or when reproducibility of the aggregation is a concern. To deal with these problems, in this work we design a flexible programmable switch by using as a building block PsPIN, a RISC-V architecture implementing the sPIN programming model. We then design, model, and analyze different algorithms for executing the aggregation on this architecture, showing performance improvements compared to state-of-the-art approaches.

The interconnect is one of the most critical components in large scale computing systems, and its impact on the performance of applications is going to increase with the system size. In this paper, we will describe SLINGSHOT, an interconnection network for large scale computing systems. SLINGSHOT is based on high-radix switches, which allow building exascale and hyper-scale datacenters networks with at most three switch-to-switch hops. Moreover, SLINGSHOT provides efficient adaptive routing and congestion control algorithms, and highly tunable traffic classes. SLINGSHOT uses an optimized Ethernet protocol, which allows it to be interoperable with standard Ethernet devices while providing high performance to HPC applications. We analyze the extent to which SLINGSHOT provides these features, evaluating it on microbenchmarks and on several applications from the datacenter and AI worlds, as well as on HPC applications. We find that applications running on SLINGSHOT are less affected by congestion compared to previous generation networks.

System noise can negatively impact the performance of HPC systems, and the interconnection network is one of the main factors contributing to this problem. To mitigate this effect, adaptive routing sends packets on non-minimal paths if they are less congested. However, while this may mitigate interference caused by congestion, it also generates more traffic since packets traverse additional hops, causing in turn congestion on other applications and on the application itself. In this paper, we first describe how to estimate network noise. By following these guidelines, we show how noise can be reduced by using routing algorithms which select minimal paths with a higher probability. We exploit this knowledge to design an algorithm which changes the probability of selecting minimal paths according to the application characteristics. We validate our solution on microbenchmarks and real-world applications on two systems relying on a Dragonfly interconnection network, showing noise reduction and performance improvement.

Summary Continuous streaming computations are usually composed of different modules, exchanging data through shared message queues. The selection of the algorithm used to access such queues (ie, theconcurrency control) is a critical aspect both for performance and power consumption. In this paper, we describe the design of automatic concurrency control algorithm for implementing power‐efficient communications on shared‐memory multicores. The algorithm automatically switches between nonblocking and blocking concurrency protocols, getting the best from the two worlds, ie, obtaining the same throughput offered by the nonblocking implementation and the same power efficiency of the blocking concurrency protocol. We demonstrate the effectiveness of our approach using two micro‐benchmarks and two real streaming applications.

This paper studies k-plexes, a well known pseudo-clique model for network communities. In a k-plex, each node can miss at most k-1 links. Our goal is to detect large communities in today's real-world graphs which can have hundreds of millions of edges. While many have tried, this task has been elusive so far due to its computationally challenging nature: k-plexes and other pseudo-cliques are harder to find and more numerous than cliques, a well known hard problem. We present D2K, which is the first algorithm able to find large k-plexes of very large graphs in just a few minutes. The good performance of our algorithm follows from a combination of graph-theoretical concepts, careful algorithm engineering and a high-performance implementation. In particular, we exploit the low degeneracy of real-world graphs, and the fact that large enough k-plexes have diameter 2. We validate a sequential and a parallel/distributed implementation of D2K on real graphs with up to half a billion edges.

In this work, we consider the C++ Actor Framework (CAF), a recent proposal that revamped the interest in building concurrent and distributed applications using the actor programming model in C++. CAF has been optimized for high-throughput computing, whereas message latency between actors is greatly influenced by the message data rate: at low and moderate rates the latency is higher than at high data rates. To this end, we propose a modification of the polling strategies in the work-stealing CAF scheduler, which can reduce message latency at low and moderate data rates up to two orders of magnitude without compromising the overall throughput and message latency at maximum pressure. The technique proposed uses a lightweight event notification protocol that is general enough to be used used to optimize the runtime of other frameworks experiencing similar issues.

Power-aware computing is gaining an increasing attention both in academic and industrial settings. The problem of guaranteeing a given QoS requirement (either in terms of performance or power consumption) can be faced by selecting and dynamically adapting the amount of physical and logical resources used by the application. In this study, we considered standard multicore platforms by taking as a reference approaches for power-aware computing two well-known dynamic reconfiguration techniques: Concurrency Throttling and Thread Packing. Furthermore, we also studied the impact of using simultaneous multithreading (e.g., Intel's HyperThreading) in both techniques. In this work, leveraging on the applications of the PARSEC benchmark suite, we evaluate these techniques by considering performance-power trade-offs, resource efficiency, predictability and required programming effort. The results show that, according to the comparison criteria, these techniques complement each other.

High-level parallel programming is an active research topic aimed at promoting parallel programming methodologies that provide the programmer with high-level abstractions to develop complex parallel software with reduced time to solution. Pattern-based parallel programming is based on a set of composable and customizable parallel patterns used as basic building blocks in parallel applications. In recent years, a considerable effort has been made in empowering this programming model with features able to overcome shortcomings of early approaches concerning flexibility and performance. In this article, we demonstrate that the approach is flexible and efficient enough by applying it on 12 out of 13 PARSEC applications. Our analysis, conducted on three different multicore architectures, demonstrates that pattern-based parallel programming has reached a good level of maturity, providing comparable results in terms of performance with respect to both other parallel programming methodologies based on pragma-based annotations (i.e., Open mp and O mp S s ) and native implementations (i.e., P threads ). Regarding the programming effort, we also demonstrate a considerable reduction in lines of code and code churn compared to P threads and comparable results with respect to other existing implementations.

Current architectures provide many control knobs for the reduction of power consumption of applications, like reducing the number of used cores or scaling down their frequency. However, choosing the right values for these knobs in order to satisfy requirements on performance and/or power consumption is a complex task and trying all the possible combinations of these values is an unfeasible solution since it would require too much time. For this reasons, there is the need for techniques that allow an accurate estimation of the performance and power consumption of an application when a specific configuration of the control knobs values is used. Usually, this is done by executing the application with different configurations and by using these information to predict its behaviour when the values of the knobs are changed. However, since this is a time consuming process, we would like to execute the application in the fewest number of configurations possible. In this work, we consider as control knobs the number of cores used by the application and the frequency of these cores. We show that on most Parsec benchmark programs, by executing the application in 1% of the total possible configurations and by applying a multiple linear regression model we are able to achieve an average accuracy of 96% in predicting its execution time and power consumption in all the other possible knobs combinations.

In current computing systems, many applications require guarantees on their maximum power consumption to not exceed the available power budget. On the other hand, for some applications, it could be possible to decrease their performance, yet maintain an acceptable level, in order to reduce their power consumption. To provide such guarantees, a possible solution consists in changing the number of cores assigned to the application, their clock frequency, and the placement of application threads over the cores. However, power consumption and performance have different trends depending on the application considered and on its input. Finding a configuration of resources satisfying user requirements is, in the general case, a challenging task. In this article, we propose Nornir, an algorithm to automatically derive, without relying on historical data about previous executions, performance and power consumption models of an application in different configurations. By using these models, we are able to select a close-to-optimal configuration for the given user requirement, either performance or power consumption. The configuration of the application will be changed on-the-fly throughout the execution to adapt to workload fluctuations, external interferences, and/or application’s phase changes. We validate the algorithm by simulating it over the applications of the Parsecbenchmark suit. Then, we implement our algorithm and we analyse its accuracy and overhead over some of these applications on a real execution environment. Eventually, we compare the quality of our proposal with that of the optimal algorithm and of some state-of-the-art solutions.

Determining the right amount of resources needed for a given computation is a critical problem. In many cases, computing systems are configured to use an amount of resources to manage high load peaks even though this cause energy waste when the resources are not fully utilised. To avoid this problem, adaptive approaches are used to dynamically increase/decrease computational resources depending on the real needs. A different approach based on Dynamic Voltage and Frequency Scaling (DVFS) is emerging as a possible alternative solution to reduce energy consumption of idle CPUs by lowering their frequencies. In this work, we propose to tackle the problem in stream parallel computations by using both the classic adaptivity concepts and the possibility provided by modern CPUs to dynamically change their frequency. We validate our approach showing a real network application that performs Deep Packet Inspection over network traffic. We are able to manage bandwidth changing over time, guaranteeing minimal packet loss during reconfiguration and minimal energy consumption.