ACM SIGCOMM 2022 Program

Abstract: As deep reinforcement learning (RL) showcases its strengths in networking, its pitfalls are also coming to the public's attention. Training on a wide range of network environments leads to suboptimal performance, whereas training on a narrow distribution of environments results in poor generalization. This work presents Genet, a new training framework for learning better RL-based network adaptation algorithms. Genet is built on curriculum learning, which has proved effective against similar issues in other RL applications. At a high level, curriculum learning gradually feeds more "difficult" environments to the training rather than choosing them uniformly at random. However, applying curriculum learning in networking is nontrivial since the "difficulty" of a network environment is unknown. Our insight is to leverage traditional rule-based (non-RL) baselines: If the current RL model performs significantly worse in a network environment than the rule-based baselines, then further training it in this environment tends to bring substantial improvement. Genet automatically searches for such environments and iteratively promotes them to training. Three case studies---adaptive video streaming, congestion control, and load balancing---demonstrate that Genet produces RL policies that outperform both regularly trained RL policies and traditional baselines.

Abstract: Adaptive neural networks (NN) have been used to optimize OS kernel datapath functions because they can achieve superior performance under changing environments. However, how to deploy these NNs remains a challenge. One approach is to deploy these adaptive NNs in the userspace. However, such userspace deployments suffer from either high cross-space communication overhead or low responsiveness, significantly compromising the function performance. On the other hand, pure kernel-space deployments also incur a large performance degradation because the computation logic of model tuning algorithm is typically complex, interfering with the performance of normal datapath execution. This paper presents LiteFlow, a hybrid solution to build high-performance adaptive NNs for kernel datapath. At its core, LiteFlow decouples the control path of adaptive NNs into: (1) a kernel-space fast path for efficient model inference, and (2) a userspace slow path for effective model tuning. We have implemented LiteFlow with Linux kernel datapath and evaluated it with three popular datapath functions including congestion control, flow scheduling, and load balancing. Compared to prior works, LiteFlow achieves 44.4% better goodput for congestion control, and improves the completion time for long flows by 33.7% and 56.7% for flow scheduling and load balancing, respectively.

Abstract: Training Deep Learning (DL) model requires multiple resource types, including CPUs, GPUs, storage IO, and network IO. Advancements in DL have produced a wide spectrum of models that have diverse usage patterns on different resource types. Existing DL schedulers focus on only GPU allocation, while missing the opportunity of packing jobs along multiple resource types. We present Muri, a multi-resource cluster scheduler for DL workloads. Muri exploits multi-resource interleaving of DL training jobs to achieve high resource utilization and reduce job completion time (JCT). DL jobs have a unique staged, iterative computation pattern. In contrast to multi-resource schedulers for big data workloads that pack jobs in the space dimension, Muri leverages this unique pattern to interleave jobs on the same set of resources in the time dimension. Muri adapts Blossom algorithm to find the perfect grouping plan for single-GPU jobs with two resource types, and generalizes the algorithm to handle multi-GPU jobs with more than two types. We build a prototype of Muri and integrate it with PyTorch. Experiments on a cluster with 64 GPUs demonstrate that Muri improves the average JCT by up to 3.6× and the makespan by up to 1.6× over existing DL schedulers.

Abstract: Network simulators are an essential tool for network operators, and can assist important tasks such as capacity planning, topology design, and parameter tuning. Popular simulators are all based on discrete event simulation, and their performance does not scale with the size of modern networks. Recently, deep-learning-based techniques are introduced to solve the scalability problem, but, as we show with experiments, they have poor visibility in their simulation results, and cannot generalize to diverse scenarios. In this work, we combine scalable and generalized continuous simulation techniques with discrete event simulation to achieve high scalability, while providing packet-level visibility. We start from a solid queueing-theoretic modeling of modern networks, and carefully identify the mathematically-intractable or computationally-expensive parts, only which are then modeled using deep neural networks (DNN). Dubbed DeepQueueNet, our approach combines prior knowledge of networks, and supports arbitrary topology and device traffic management mechanisms (given sufficient training data). Our extensive experiments show that DeepQueueNet achieves near-linear speedup in the number of GPUs, and its estimation accuracy for average and 99th percentile round-trip time outperforms existing end-to-end DNN-based performance estimators in all scenarios.

Abstract: We explore the feasibility of using Generative Adversarial Networks (GANs) to automatically learn generative models to generate synthetic packet- and flow header traces for networking tasks (e.g., telemetry, anomaly detection, provisioning). We identify key fidelity, scalability, and privacy challenges and tradeoffs in existing GAN-based approaches. By synthesizing domain-specific insights with recent advances in machine learning and privacy, we identify design choices to tackle these challenges. Building on these insights, we develop an end-to-end framework, NetShare. We evaluate NetShare on six diverse packet header traces and find that: (1) across all distributional metrics and traces, it achieves 46% more accuracy than baselines and (2) it meets users' requirements of downstream tasks in evaluating accuracy and rank ordering of candidate approaches.

Abstract: Round-trip time (RTT) is a central metric that influences end-user QoE and can expose traffic-interception attacks. Many popular RTT monitoring techniques either send active probes (that do not capture application-level RTTs) or passively monitor only the TCP handshake (which can be inaccurate, especially for long-lived flows). High-speed programmable switches present a unique opportunity to monitor the RTTs continuously and react in real time to improve performance and security. In this paper, we present Dart, an inline, real-time, and continuous RTT measurement system that can enable automated detection of network events and adapt (e.g., routing, scheduling, marking, or dropping traffic) inside the network. However, designing Dart is fraught with challenges, due to the idiosyncrasies of the TCP protocol and the resource constraints in high-speed switches. Dart overcomes these challenges by strategically limiting the tracking of packets to only those that can generate useful RTT samples, and by identifying the synergy between per-flow state and per-packet state for efficient memory use. We present a P4 prototype of Dart for the Tofino switch, as well our experiments on a campus testbed and simulations using anonymized campus traces. Dart, running in real time and with limited data-plane memory, is able to collect 99% of the RTT samples of an offline, software baseline---a variant of the popular tcptrace tool that has access to unlimited memory.

Abstract: Network measurement is important to data center operators. Most existing efforts focus on developing new implementation schemes for measurement tasks. Little attention is paid to on-the-fly task reconfiguration. Due to resource constraints, it is impossible to configure all needed tasks at start-up and dynamically turn on/of them. To support real-time reconfiguration of many different tasks, a key observation is that it is unnecessary to bind a task and its implementation at the compilation phase. We design FlyMon, the first sketch-based measurement system that can make on-the-fly reconfigurations on a large set of measurement tasks. FlyMon introduces the concept of Composable Measurement Units (CMUs), which are general operation units that support reconfigurable implementation for measurement tasks combined from different flow keys and flow attributes. FlyMon maps the design of CMUs to programmable switches' data planes so that the number of compacted CMUs can be maximized. FlyMon also provides dynamic memory management. We prototype FlyMon on Tofino and currently enable four frequently used flow attributes. Each CMU Group (with 3 CMUs) can concurrently perform up to 96 isolated measurement tasks with less than 8.3% hardware resources. The tasks can be deployed with configurable memory size at the millisecond level. By cross-stacking, FlyMon can deploy up to 27 CMUs in one pipeline of Tofino.

Abstract: Internet-wide scanning is commonly used to understand the topology and security of the Internet. However, IPv4 Internet scans have been limited to scanning only a subset of services---exhaustively scanning all IPv4 services is too costly and no existing bandwidth-saving frameworks are designed to scan IPv4 addresses across all ports. In this work we introduce GPS, a system that efficiently discovers Internet services across all ports. GPS runs a predictive framework that learns from extremely small sample sizes and is highly parallelizable, allowing it to quickly find patterns between services across all 65K ports and a myriad of features. GPS computes service predictions in 13 minutes (four orders of magnitude faster than prior work) and finds 92.5% of services across all ports with 131× less bandwidth, and 204× more precision, compared to exhaustive scanning. GPS is the first work to show that, given at least two responsive IP addresses on a port to train from, predicting the majority of services across all ports is possible and practical.

Abstract: When diagnosing performance anomalies, it is often useful to reason about why a packet experienced the queuing that it did. To that end, we observe that queuing is both a result of historical effects and the current state of the network. Further, both factors involve short and long timescales by nature. Existing work fails to provide insight that satisfies all of these needs. This paper presents PrintQueue, a practical data-plane monitoring system for tracking the provenance of packet-level delays at both small and large timescales. We propose a set of metrics for describing 'congestion regimes' and present a set of novel data-plane data structures that accurately track those metrics over arbitrary time spans. We implement PrintQueue on a Tofino switch and evaluate it with multiple network traces. Our evaluation shows that the accuracy of PrintQueue is up to 3× times higher while the overhead is 20× times smaller than existing work.

Abstract: As network speeds have increased to over 100 Gbps, operators and researchers have lost the ability to easily ask complex questions of reassembled and parsed network traffic. In this paper, we introduce Retina, a software framework that lets users analyze over 100 Gbps of real-world traffic on a single server with no specialized hardware. Retina supports running arbitrary user-defined analysis functions on a wide variety of extensible data representations ranging from raw packets to parsed application-layer handshakes. We introduce a novel filtering mechanism and subscription interface to safely and efficiently process high-speed traffic. Under the hood, Retina implements an efficient data pipeline that strategically discards unneeded traffic and defers expensive processing operations to preserve computation for complex analyses. We present the framework architecture, evaluate its performance on production traffic, and explore several applications. Our experiments show that Retina is capable of running sophisticated analyses at over 100 Gbps on a single commodity server and can support 5--100× higher traffic rates than existing solutions, dramatically reducing the effort to complete investigations on real-world networks.

Abstract: Since its inception, underwater digital acoustic communication has required custom hardware that neither has the economies of scale nor is pervasive. We present the first acoustic system that brings underwater messaging capabilities to existing mobile devices like smartphones and smart watches. Our software-only solution leverages audio sensors, i.e., microphones and speakers, ubiquitous in today's devices to enable acoustic underwater communication between mobile devices. To achieve this, we design a communication system that in real-time adapts to differences in frequency responses across mobile devices, changes in multipath and noise levels at different locations and dynamic channel changes due to mobility. We evaluate our system in six different real-world underwater environments with depths of 2--15 m in the presence of boats, ships and people fishing and kayaking. Our results show that our system can in real-time adapt its frequency band and achieve bit rates of 100 bps to 1.8 kbps and a range of 30 m. By using a lower bit rate of 10--20 bps, we can further increase the range to 100 m. As smartphones and watches are increasingly being used in underwater scenarios, our software-based approach has the potential to make underwater messaging capabilities widely available to anyone with a mobile device. Project page with open-source code and data can be found here: https://underwatermessaging.cs.washington.edu/.

Abstract: Given the increasing number of building collapse tragedies nowadays (e.g., Florida condo collapse), people gradually recognize that long-term and persistent structural health monitoring (SHM) becomes indispensable for civilian buildings. However, current SHM techniques suffer from high cost and deployment difficulty caused by the wired connection. Traditional wireless sensor networks fail to serve in-concrete communication for SHM because of the complexity of battery replacement and the concrete Faraday cage. In this work, we collaborate with experts from civil engineering to create a type of promising self-sensing concrete by introducing a novel functional filler, called EcoCapsule- a battery-free and miniature piezoelectric backscatter node. We overcome the fundamental challenges in in-concrete energy harvesting and wireless communication to achieve SHM via EcoCapsules. We prototype EcoCapsules and mix them with other raw materials (such as cement, sand, water, etc) to cast the self-sensing concrete, into which EcoCapsules are implanted permanently. We tested EcoCapsules regarding real-world buildings comprehensively. Our results demonstrate single link throughputs of up to 13 kbps and power-up ranges of up to 6 m. Finally, we demonstrate a long-term pilot study on the structural health monitoring of a real-life footbridge.

Abstract: Backscatter communication enables miniature, batteryless, and low-cost wireless sensors. Since electromagnetic waves are strongly attenuated in several scenarios, backscatter communication in metals via acoustic waves can leverage various applications, e. g., in structural health monitoring. When backscattering, the Tag has little control over the modulation it performs on the carrier wave. Therefore, existing approaches commonly employ differential binary modulation schemes, limiting the achievable data rates. To overcome this limitation, we derive a channel model that accurately describes the modulation in an acoustic backscatter channel---as, e. g., found in steel beams---and leverage it to achieve higher-order load modulation. We present an open-source Reader and Tag pair prototype based on COTS components that we have developed for communication and on-the-fly channel characterization. We explore the influence of various parameters on communication performance on different channels. Moreover, (i) we are the first to demonstrate that acoustic backscatter is feasible in guided-wave channels, covering up to 3 meters, and that (ii) our modulation scheme achieves up to 211% higher data rates than binary modulation schemes, and (iii) provides reliable communication through channel coding.

Abstract: The advent of radio sensing that works through walls & obstacles challenges the notion of indoor privacy. An eavesdropper can deploy such sensing to snoop on their neighbors and a smart sensor embedded with such sensing capabilities can perform large scale behavioral and health data mining. We present RF-Protect, a new framework that enables privacy by injecting fake humans in the sensed data. RF-Protect consists of a novel hardware reflector design that modifies radio waves to create reflections at arbitrary locations in the environment and a new generative mechanism to create realistic human trajectories. RF-Protect's design doesn't require any high bandwidth hardware or physical motion. We implement RF-Protect using commodity hardware and validate its ability to generate fake human trajectories.

Abstract: The ultimate goal of virtual reality (VR) is to create an experience indistinguishable from actual reality. To provide such a "life-like" experience, (i) the VR headset (VRH) should be wireless so that the user can move around freely, and (ii) the wireless link, connecting the VRH to a high-performance renderer, should support high data rates (tens to hundreds of Gbps). Industry is already pushing towards such wireless VRHs; however, these wireless links can only support a few Gbps rates. In general, current radio-frequency (RF) links (including mmWave) are not able to provide desired data rates. In this paper, we build a system, we call Cyclops, which uses free-space optical (FSO) technology to create a high-bandwidth VR wireless link. FSO links are capable of very high data rates (up to Tbps) due to the high frequencies of light waves and narrow beams. The main challenges in developing an effective FSO link are: (i) designing a link with sufficient movement tolerance, and (ii) developing a viable tracking and pointing (TP) mechanism which maintains the link while the VRH moves. As traditional TP approaches seem infeasible in our context, we develop a novel TP approach based on learning techniques, leveraging the VRH's inbuilt tracking system. We build robust 10 Gbps and 25Gbps link prototypes from commodity components, demonstrate their viability for expected movement speeds of a VRH, and show that, with certain custom-built components, we can support much higher movement speeds and bandwidths.

Abstract: n multi-tenant data centers, each tenant desires reassuring predictability from the virtual network fabric - bandwidth guarantee, work conservation, and bounded tail latency. Achieving these goals simultaneously relies on rapid and precise traffic admission. However, the slow convergence (tens of milliseconds) of prior works can hardly satisfy the increasingly rigorous performance demand under dynamic traffic patterns. Further, state-of-the-art load balance schemes are all guarantee-agnostic and bring great risks on breaking bandwidth guarantee, which is overlooked in prior works. In this paper, we propose μFab, a predictable virtual fabric solution which can (1) explicitly select proper paths for all flows and (2) converge to ideal bandwidth allocation at sub-millisecond timescales. The core idea of μFab is to leverage the programmable data plane to build a fusion of an active edge (e.g., NIC) and an informative core (e.g., switch), where the core sends link status and tenant information to the edge via telemetry to help the latter make a timely and accurate decision on path selection and traffic admission. We fully implement μFab with commodity SmartNICs and programmable switches. Evaluations show that μFab can keep minimum bandwidth guarantee with high bandwidth utilization and near-optimal transmission latency in various network situations with limited probing bandwidth overhead. Application-level experiments, e.g., compute and storage scenarios, show that μFab can improve QPS by 2.5× and cut tail latency by more than 21× compared to the alternatives.

Abstract: This paper describes Trio, a programmable chipset used in Juniper Networks' MX-series routers and switches. Trio's architecture is based on a multi-threaded programmable packet processing engine and a hierarchy of high-capacity memory systems, making it fundamentally different from pipeline-based architectures. Trio gracefully handles non-homogeneous packet processing rates for a wide range of networking use cases and protocols, making it an ideal platform for emerging in-network applications. We begin by describing the Trio chipset's fundamental building blocks, including its multi-threaded Packet Forwarding and Packet Processing Engines. We then discuss Trio's programming language, called Microcode. To showcase Trio's flexible Microcode-based programming environment, we describe two use cases. First, we demonstrate Trio's ability to perform in-network aggregation for distributed machine learning. Second, we propose and design an in-network straggler mitigation technique using Trio's timer threads. We prototype both use cases on a testbed using three real DNN models (ResNet50, DenseNet161, and VGG11) to demonstrate Trio's ability to mitigate stragglers while performing in-network aggregation. Our evaluations show that when stragglers occur in the cluster, Trio outperforms today's pipeline-based solutions by up to 1.8x.

Abstract: The ability to filter entries in the data plane from a set or table of resources (e.g., network paths, servers, switch ports) based on multi-dimensional policies over stateful resource-specific metrics (e.g., filter paths with utilization < 0.6 and latency < 3us) is critical for several key network functions, such as performance-aware routing, resource-aware load balancing, network diagnosis, security and firewall. However, current generation of programmable switches do not support table-wide stateful filtering at line rate. We present Thanos, which augments the existing programmable switch pipeline with support for programmable multi-dimensional filtering over a set of resources. Thanos seamlessly integrates with multi-terabit programmable switch pipelines at nominal chip area overhead. Our evaluation, based on an FPGA prototype and a simulator, shows that policies expressed in Thanos can improve the performance of key network functions by up to 1.7× compared to state-of-the-art.

Abstract: Given the clock rate of a single packet processing pipeline has saturated due to slowdown in transistor scaling, today's programmable switches employ multiple parallel pipelines to meet high packet processing rates. However, parallel processing poses a challenge for stateful packet processing, where it becomes hard to guarantee functional correctness while maintaining line rate processing. This paper presents the design and implementation of MP5, which is a new switch architecture, compiler, and runtime for multi-pipelined programmable switches that is functionally equivalent to a logical single pipelined switch while also processing packets close to the ideal processing rate, for all packet processing programs.

Abstract: Avoiding packet loss is crucial for ISPs. Unfortunately, malfunctioning hardware at ISPs can cause long-lasting packet drops, also known as gray failures, which are undetectable by existing monitoring tools. In this paper, we describe the design and implementation of FANcY, an ISP-targeted system that detects and localizes gray failures quickly and accurately. FANcY complements previous monitoring approaches, which are mainly tailored for low-delay networks such as data center networks and do not work at ISP scale. We experimentally confirm FANcY's capability to accurately detect gray failures in seconds, as long as only tiny fractions of traffic experience losses. We also implement FANcY in an Intel Tofino switch, demonstrating how it enables fine-grained fast rerouting.