Multi-hop all-reduce is the de facto backbone of large model training. As the training scale increases, the network often becomes a bottleneck, motivating the reduction of the volume of transmitted data. Accordingly, recent systems have demonstrated significant acceleration of the training process using gradient quantization. However, these systems are not optimized for multi-hop aggregation, where entries are partially summed multiple times along their aggregation topology.

Mixture-of-Experts (MoE) models improve scaling by activating a small number of experts per token. However, the combined memory requirements of all experts may exceed the GPU's available high-bandwidth memory (HBM) during inference. Inference frameworks such as vLLM and HuggingFace address the problem by offloading experts to CPU memory and moving them to the GPU's HBM as needed. While enabling inference of large models with limited HBM, this CPU-GPU traffic overhead slows down token generation.

Disaggregated Large Language Model (LLM) inference decouples the compute-intensive prefill stage from the memory-intensive decode stage, allowing low-end, compute-focused GPUs for prefill and high-end, memory-rich GPUs for decode, which reduces cost while maintaining high throughput. However, transmitting Key-Value (KV) data between the two stages can be a bottleneck, especially for long prompts. Additionally, the computational overhead in the two stages is key for optimizing Job Completion Time (JCT), and KV data size can become prohibitive for long prompts and sequences. Existing KV quantization methods can alleviate transmission and memory bottlenecks, but they introduce significant dequantization overhead, exacerbating the computation time.

Distributed training of ML models generates significant network traffic when exchanging gradients and is sensitive to packet drops and retransmission caused by congestion when other traffic is sharing the network. While training processes can tolerate gradient compression to reduce traffic volume, it does not eliminate queue buildup in shallow-buffer switches, causing high latency and extended training time.

Gradient aggregation has long been identified as a major bottleneck in today's large-scale distributed machine learning training systems. One promising solution to mitigate such bottlenecks is gradient compression, directly reducing communicated gradient data volume. However, in practice, many gradient compression schemes do not achieve acceleration of the training process while also preserving accuracy.

Distributed Mean Estimation (DME), in which $n$ clients communicate vectors to a parameter server that estimates their average, is a fundamental building block in communication-efficient federated learning. In this paper, we improve on previous DME techniques that achieve the optimal $O(1/n)$ Normalized Mean Squared Error (NMSE) guarantee by asymptotically improving the complexity for either encoding or decoding (or both). To achieve this, we formalize the problem in a novel way that allows us to use off-the-shelf mathematical solvers to design the quantization. Using various datasets and training tasks, we demonstrate how QUIC-FL achieves state of the art accuracy with faster encoding and decoding times compared to other DME methods.

Quantization is a fundamental optimization for many machine learning (ML) use cases, including compressing gradients, model weights and activations, and datasets. The most accurate form of quantization is adaptive, where the error is minimized with respect to a given input rather than optimizing for the worst case. However, optimal adaptive quantization methods are considered infeasible in terms of both their runtime and memory requirements. We revisit the Adaptive Stochastic Quantization (ASQ) problem and present algorithms that find optimal solutions with asymptotically improved time and space complexities. Our experiments indicate that our algorithms may open the door to using ASQ more extensively in a variety of ML applications. We also present an even faster approximation algorithm for quantizing large inputs on the fly.

The soaring use of machine learning leads to increasing processing demands. As data volume keeps growing, providing classification services with good machine learning performance, high throughput, low latency, and minimal equipment overheads becomes a challenge. Offloading machine learning tasks to network switches can be a scalable solution to this problem, providing high throughput and low latency. However, network devices are resource constrained, and lack support for machine learning functionality. In this paper, we introduce IIsy -a novel mapping tool of machine learning classification models to off-the-shelf switches. Using an efficient encoding algorithm, enables fitting a range of classification models on switches, co-existing with standard switch functionality. To overcome resource constraints, adopts a hybrid approach for ensemble models, running a small model on a switch and a large model on the backend. The evaluation shows that achieves near-optimal classification results, within minimum resource overheads, and while reducing the load on the backend by 70% for data-intensive use cases.

Many compression techniques have been proposed to reduce the communication overhead of Federated Learning training procedures. However, these are typically designed for compressing model updates, which are expected to decay throughout training. As a result, such methods are inapplicable to downlink (i.e., from the parameter server to clients) compression in the cross-device setting, where heterogeneous clients *may appear only once* during training and thus must download the model parameters. Accordingly, we propose DoCoFL -- a new framework for downlink compression in the cross-device setting. Importantly, DoCoFL can be seamlessly combined with many uplink compression schemes, rendering it suitable for bi-directional compression. Through extensive evaluation, we show that DoCoFL offers significant bi-directional bandwidth reduction while achieving competitive accuracy to that of a baseline without any compression.

With the rapid increase in the size and volume of cloud services and data centers, architectures with multiple job dispatchers are quickly becoming the norm. Load balancing is a key element of such systems. Nevertheless, current solutions to load balancing in such systems admit a paradoxical behavior in which more accurate information regarding server queue lengths degrades performance due to herding and detrimental incast effects. Indeed, both in theory and in practice, there is a common doubt regarding the value of information in the context of multi-dispatcher load balancing. As a result, both researchers and system designers resort to more straightforward solutions, such as the power-of-two-choices to avoid worst-case scenarios, potentially sacrificing overall resource utilization and system performance. A principal focus of our investigation concerns the value of information about queue lengths in the multi-dispatcher setting. We argue that, at its core, load balancing with multiple dispatchers is a distributed computing task. In that light, we propose a new job dispatching approach, called Tidal Water Filling, which addresses the distributed nature of the system. Specifically, by incorporating the existence of other dispatchers into the decision-making process, our protocols outperform previous solutions in many scenarios. In particular, when the dispatchers have complete and accurate information regarding the server queues, our policies significantly outperform all existing solutions.

Distributed Mean Estimation (DME) is a central building block in federated learning, where clients send local gradients to a parameter server for averaging and updating the model. Due to communication constraints, clients often use lossy compression techniques to compress the gradients, resulting in estimation inaccuracies. DME is more challenging when clients have diverse network conditions, such as constrained communication budgets and packet losses. In such settings, DME techniques often incur a significant increase in the estimation error leading to degraded learning performance. In this work, we propose a robust DME technique named EDEN that naturally handles heterogeneous communication budgets and packet losses. We derive appealing theoretical guarantees for EDEN and evaluate it empirically. Our results demonstrate that EDEN consistently improves over state-of-the-art DME techniques.

Cloud operators require timely 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 present the case for identifying heavy hitters throughsliding windows. Sliding windows are quicker and more accurate to detect new heavy hitters than current interval-based methods, but to date had no practical algorithms. Accordingly, we introduce, design, and analyze theMementofamily of sliding window algorithms for the HH and HHH problems in the single-device and network-wide settings. We use extensive evaluations to show that our single-device solutions are orders of magnitude faster than existing sliding window techniques and comparable in speed to state-of-the-art non-windowed sampling based technique. 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.

Hash-based stateful load-balancers employ connection tracking to avoid per-connection-consistency (PCC) violations that lead to broken connections. In this paper, we propose Just Enough Tracking (JET), a new algorithmic framework that significantly reduces the size of the connection tracking tables for hash-based stateful load-balancers without increasing PCC violations.

Direct memory access (DMA) renders a system vulnerable to DMA attacks, in which I/O devices access memory regions not intended for their use. Hardware input-output memory management units (IOMMU) can be used to provide protection. However, an IOMMU cannot prevent all DMA attacks because it only restricts DMA at page-level granularity, leading to sub-page vulnerabilities.

We consider the fundamental problem of communicating an estimate of a real number $x\in[0,1]$ using a single bit. A sender that knows $x$ chooses a value $X\in\set{0,1}$ to transmit. In turn, a receiver estimates $x$ based on the value of $X$. We consider both the biased and unbiased estimation problems and aim to minimize the cost. For the biased case, the cost is the worst-case (over the choice of $x$) expected squared error, which coincides with the variance if the algorithm is required to be unbiased. We first overview common biased and unbiased estimation approaches and prove their optimality when no shared randomness is allowed. We then show how a small amount of shared randomness, which can be as low as a single bit, reduces the cost in both cases. Specifically, we derive lower bounds on the cost attainable by any algorithm with unrestricted use of shared randomness and propose near-optimal solutions that use a small number of shared random bits. Finally, we discuss open problems and future directions.

Counters are the fundamental building block of many data sketching schemes, which hash items to a small number of counters and account for collisions to provide good approximations for frequencies and other measures. Most existing methods rely on fixed-size counters, which may be wasteful in terms of space, as counters must be large enough to eliminate any risk of overflow. Instead, some solutions use small, fixed-size counters that may overflow into secondary structures.This paper takes a different approach. We propose a simple and general method called SALSA for dynamic re-sizing of counters, and show its effectiveness. SALSA starts with small counters, and overflowing counters simply merge with their neighbors. SALSA can thereby allow more counters for a given space, expanding them as necessary to represent large numbers. Our evaluation demonstrates that, at the cost of a small overhead for its merging logic, SALSA significantly improves the accuracy of popular schemes (such as Count-Min Sketch and Count Sketch) over a variety of tasks. Our code is released as open source [1].

We consider the problem where $n$ clients transmit $d$-dimensional real-valued vectors using $d(1+o(1))$ bits each, in a manner that allows the receiver to approximately reconstruct their mean. Such compression problems naturally arise in distributed and federated learning. We provide novel mathematical results and derive computationally efficient algorithms that are more accurate than previous compression techniques. We evaluate our methods on a collection of distributed and federated learning tasks, using a variety of datasets, and show a consistent improvement over the state of the art.

Current-day data centers and high-volume cloud services employ a broad set of heterogeneous servers. In such settings, client requests typically arrive at multiple entry points, and dispatching them to servers is an urgent distributed systems problem. This paper presents an efficient solution to the load balancing problem in such systems that improves on and overcomes problems of previous solutions. The load balancing problem is formulated as a stochastic optimization problem, and an efficient algorithmic solution is obtained based on a subtle mathematical analysis of the problem. Finally, extensive evaluation of the solution on simulated data shows that it outperforms previous solutions. Moreover, the resulting dispatching policy can be computed very efficiently, making the solution practically viable.

Consistent hashing is a central building block in many networking applications, such as maintaining connection affinity of TCP flows. However, current consistent hashing solutions do not ensure full consistency under arbitrary changes or scale poorly in terms of memory footprint, update time and key lookup complexity. We present AnchorHash, a scalable and fully-consistent hashing algorithm. AnchorHash achieves high key lookup rate, low memory footprint and low update time. We formally establish its strong theoretical guarantees, and present an advanced implementation with a memory footprint of only a few bytes per resource. Moreover, evaluations indicate that AnchorHash scales on a single core to 100 million resources while still achieving a key lookup rate of more than 15 million keys per second.

Counters are a fundamental building block for networking applications such as load balancing, traffic engineering, and intrusion detection, which require estimating flow sizes and identifying heavy hitter flows. Existing works suggest replacing counters with shorter multiplicative error estimators that improve the accuracy by fitting more of them within a given space. However, such estimators impose a computational overhead that degrades the measurement throughput. Instead, we propose additive error estimators, which are simpler, faster, and more accurate when used for network measurement. Our solution is rigorously analyzed and empirically evaluated against several other measurement algorithms on real Internet traces. For a given error target, we improve the speed of the uncompressed solutions by 5×-30×, and the space by up to 4×. Compared with existing state-of-the-art estimators, our solution is 9×-35× faster while being considerably more accurate.

With the rapid increase in the size and volume of cloud services and data centers, architectures with multiple job dispatchers are quickly becoming the norm. Load balancing is a key element of such systems. Nevertheless, current solutions to load balancing in such systems admit a paradoxical behavior in which more accurate information regarding server queue lengths degrades performance due to herding and detrimental incast effects. Indeed, both in theory and in practice, there is a common doubt regarding the value of information in the context of multi-dispatcher load balancing. As a result, both researchers and system designers resort to more straightforward solutions, such as the power-of-two-choices to avoid worst-case scenarios, potentially sacrificing overall resource utilization and system performance. A principal focus of our investigation concerns the value of information about queue lengths in the multi-dispatcher setting. We argue that, at its core, load balancing with multiple dispatchers is a distributed computing task. In that light, we propose a new job dispatching approach, called Tidal Water Filling, which addresses the distributed nature of the system. Specifically, by incorporating the existence of other dispatchers into the decision-making process, our protocols outperform previous solutions in many scenarios. In particular, when the dispatchers have complete and accurate information regarding the server queue lengths, our policies significantly outperform all existing solutions.

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 loadbalancers, 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.

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.

Network providers now want to enable users to define their own flow priorities, and commercial devices already implement this ability. However, it has been shown that directly applying arbitrary user-defined priorities can fundamentally destabilize a network. In this paper, we show that it is possible to apply user-defined priorities while keeping the network stable. We introduce U-BP, a scalable approach that extends backpressure-based scheduling techniques to service user-defined flow priorities and rates while maintaining throughput optimality and strong network performance. We explain how our approach relies on a dual-layer scheme with an exponential convergence to requested priorities. We further prove analytically the network stability of our solution, and show how it achieves a strong performance for high-priority flows.

We present dRMT (disaggregated Reconfigurable Match-Action Table), a new architecture for programmable switches. dRMT overcomes two important restrictions of RMT, the predominant pipeline-based architecture for programmable switches: (1) table memory is local to an RMT pipeline stage, implying that memory not used by one stage cannot be reclaimed by another, and (2) RMT is hardwired to always sequentially execute matches followed by actions as packets traverse pipeline stages. We show that these restrictions make it difficult to execute programs efficiently on RMT.

Backpressure schemes are known to stabilize stochastic networks through the use of congestion gradients in routing and resource allocation decisions. Nonetheless, these schemes share a significant drawback, namely, the delay guarantees are obtained only in terms of average values. As a result, arbitrary packets may never reach their destination due to both the starvation and last-packet problems. These problems occur because in backpressure schemes, packet scheduling needs a subsequent stream of packets to produce the required congestion gradient for scheduling. To solve these problems, we define a starvation-free stability criterion that ensures a repeated evacuation of all network queues. Then, we introduce SF-BP, the first backpressure routing and resource allocation algorithm that is starvation-free stable. We further present stronger per-queue service guarantees and provide tools to enhance weak streams. We formally prove that our algorithm ensures that all packets reach their destination for wide families of networks. Finally, we verify our results by extensive simulations using challenging topologies as well as random static and dynamic topologies.

Hybrid switching combines a high-bandwidth optical circuit switch in parallel with a low-bandwidth electronic packet switch. It presents an appealing solution for scaling datacenter architectures. Unfortunately, it does not fit many traffic patterns produced by typical datacenter applications, and in particular the skewed traffic patterns that involve highly intensive one-to-many and many-to-one communications.

New congestion control algorithms are rapidly improving datacenters by reducing latency, overcoming incast, increasing throughput and improving fairness. Ideally, the operating system in every server and virtual machine is updated to support new congestion control algorithms. However, legacy applications often cannot be upgraded to a new operating system version, which means the advances are off-limits to them. Worse, as we show, legacy applications can be squeezed out, which in the worst case prevents the entire network from adopting new algorithms.

Over the recent years we witness a massive growth of cloud usage, accelerated by new types of 'born-to-the-cloud' workloads. These new types of workloads are increasingly multi-component, dynamic and often present highly intensive communication patterns. Massive innovation of Data Center Network (DCN) technologies is required to support the demand, giving raise to new network topologies, new network control paradigms, and management models. One particularly promising technology candidate for improving the DCN efficiency is Optical Circuit Switching (OCS).