My recent research has developed principled and cross-disciplinary approaches to exploit the potential that AI offers for the entire gamut of tasks in designing and managing computer networks ranging from (i) the management plane (e.g., network configuration management, performance diagnosis and what-if analysis); (ii) the control plane (e.g., routing, traffic engineering); and (iii) the data plane (per-packet forwarding at terabits per second). My research has developed (i) custom domain-inspired ML architectures that can handle workloads (e.g., traffic demands, failure scenarios) significantly different than training, and causal biases in network trace-driven analysis; (ii) systems that combine AI-assisted formalization, and formal verification methods to audit network configurations; and (iii) algorithms for resource-efficient and reprogrammable real-time ML inference that leverage programmable networking. Below are some recent directions and representative publications.
This paper presents HARP, the first neural model for WAN TE explicitly capable of handling variations in topology (that may occur due to WAN evolution, failures or planned maintenance) including those not observed in training. HARP is designed with two principles in mind: (i) ensure invariances to natural input transformations (e.g., permutations of node ids, tunnel reordering); and (ii) align neural architecture to the optimization model. Evaluations on real-world topologies and traffic data show that Harp achieves 5-13% lower utilizations than classic optimization solvers owing to its better resilience to prediction error, while being 100X faster. Further, it achieves 34% lower utilizations on average than prior ML-based TE schemes when encountering topologies outside the training set.
Hattrick is the first ML-based TE scheme for handling multiple traffic classes, a key requirement of cloud and ISP WANs. Hattrick's design includes (i) a novel neural architecture aligned with the sequence of optimization problems in multiclass TE; and (ii) a variant of classical multitask learning methods to deal with the unique challenge of optimizing multiple metrics that have a precedence relationship. Hattrick integrates a differentiable, class-aware flow-based simulator into its training process to effectively models detrimental effects such as capacity violations that may occur under prediction error. This leverages the unique opportunity afforded by ML-based approaches (over classical optimization methods) to incorporate such simulators, thereby enhancing the robustness and reliability of TE in the face of prediction uncertainty. Hattrick outperforms Microsoft's state-of-the-art SWAN multi-class TE controller by upto 19.3% on a 99th percentile flow delivery metric.
In this paper, we present Veritas, the first framework that tackles causal reasoning for video streaming without requiring data collected through RCTs. Causal reasoning is challenging owing to the intrinsic network bandwidth acting as latent confounder, and owing to the cascaded effects that past ABR decisions have on the future. Integral to Veritas is an easy-to-interpret domain-specific ML model that allows necessary latent variables to be inferred from actual observations (chunk download times) while exploiting knowledge of TCP states (e.g., congestion window) to facilitate inference. Validations with data of real video sessions show Veritas acccurately tackles a wide range of what-if questions, often close to an ideal oracle, while not requiring RCT training data.
Elastispec is a first step towards principled management of enterprise firewall policies with informal and incomplete specifications. Elastispec makes several contributions: (i) a custom Domain Specific Language (DSL) that is sufficiently expressive to capture policies in vendor documents of popular enterprise applications and which decouples translation and auditing; (ii) a multi-step consistency-preserving LLM approach that increases translation accuracies from 61% to 96%; (iii) an approach to map the DSL to a concrete network environment by correlating multiple enterprise data sources uncovering 2.6X more application instances than a baseline limited by imperfect inventories and (iv) a methodology to audit configurations against partial specifications, generating compliance trees and identifying out-of-spec flows. Evaluations with real-world enterprise firewall configurations show Elastispec uncovers 15 configuration errors confirmed by operators, and enables comparative analysis across multiple application instances.
In this paper, we present Leo, a system for online traffic classification at multi-terabit line rates. Leo implements an online machine learning (ML) model for traffic classification, namely the decision tree, in the network switch’s data plane. Leo’s design is fast (can classify packets at switch’s line rate), scalable (can automatically select a resource-efficient design for the class of decision tree models a user wants to support), and runtime programmable (the model can be updated on-thefly without switch downtime), while achieving high model accuracy. We implement Leo on top of Intel Tofino switches. Our evaluations show that Leo is able to classify traffic at line rate with nominal latency overhead, can scale to model sizes more than twice as large as state-of-the-art data plane ML classification systems, while achieving classification accuracy on-par with an offline traffic classifier.
Existing traffic engineering (TE) schemes either (i) optimally reroute traffic on each failure through centralized mechanisms, thereby achieving high efficiencies but with poor resilience since network links may be congested during failure recovery; or (ii) proactively guarantee the network remains congestion-free during failures by allocating bandwidth conservatively and using fast local repair mechanisms. Unfortunately, these approaches carry far less traffic than the network is capable. In this paper, we developed PCF, a novel set of congestion-free routing mechanisms that bridge this gap. PCF achieves these goals by better modeling network structure, and by carefully enhancing the flexibility of network response while ensuring that the performance under failures can be tractably modeled. All of PCF’s schemes involve relatively light-weight operations on failures. Our empirical results over 21 Internet topologies show PCF sustains higher throughput than the state-of-the-art congestion-free mechnasism (from Microsoft Research) by a factor of 1.11X to 1.5X on average across the topologies, while providing a benefit of 2.6X in some cases.
This paper has developed a novel approach to formally certify and synthesize network designs that are provably congestion-free over combinatorially many scenarios that a network may operate in (e.g., traffic demands, failures) while modeling flexible real-world network response strategies. The paper does so by establishing surprising connections between optimization-theoretic techniques, and mechanisms that networks use to respond to failures. This work significantly advances existing practice which uses ad-hoc simulation-based testing, and prior theory (e.g., robust optimization, oblivious routing) which only apply to highly restrictive network response, and to the worst-case. My subsequent work in Sigmetrics 2020 was the first to show how a network could be designed to be provably congestion-free over scenarios that occur a desired percentage of time rather than just consider the worst-case, which is conservative.
This paper presented a formal approach for verifying that low-level router configurations correctly implement Class of Service (CoS) policies for differential treatment of traffic (e.g., prioritizing video over other traffic). A tool implemented using Binary Decision diagrams was successfully used to discover anomalies in 150 real-world production networks. This work was done in collaboration with AT&T Research, and was awarded a joint patent. This paper was an early and influential work in the area of network verification, and the first to consider Class of Service policies, an important domain in networking.
The tremendous success of Internet video has led to a growing interest in new forms of video content such as 360° video. Encoding and streaming the entire 360° view to the client is prohibitively expensive, and can consume 5-6X the bandwidth needed to stream only the relevant portions of the video to the user. While streaming only the user's viewport can help, it is challenging to accurately predict the viewport, and mispredictions may hurt interactive experience. In this paper, we present Dragonfly, a new 360° system that preserves interactive experience by avoiding playback stalls while maintaining high perceptual quality. A central contribution of Dragonfly is proactive skipping, where some viewport tiles can be prudently and deliberately skipped to optimize overall experience. Using a user study with 26 users and emulation-based experiments we show that Dragonfly has higher quality, and lower overheads, than state-of-the-art 360° streaming approaches.
Ensuring high quality Internet video delivery requires ABR algorithms whose performance depends on network-sensitive parameters. This paper presents Oboe, a system that auto-tunes ABR algorithm parameters to variable and heterogeneous network conditions. Oboe significantly improves state-of-the-art approaches including MPC (from Carnegie Mellon), and BOLA (an algorithm adopted by dash.js, a popular open-source video streaming software). Further, Oboe betters Pensieve, a recent reinforcement learning based ABR from MIT, by 24% on a composite video delivery metric by better specializing ABR behavior across network states. The paper has been released with bandwidth traces of video streaming sessions, and has already been cited 480 times since publication.
The conventional wisdom throughout the 1990's was that router level support (IP Multicast) was necessary to efficiently support multicast functionality. Despite tens of hundreds of publications through the 1990's, extensive Internet standardization efforts, and implementation by router vendors such as Cisco, IP Multicast saw limited success. The above paper made the argument that an end system overlay approach with no router support could efficiently support multicast functionality. The paper opened a new area of research in peer-to-peer multicast, won the ACM Sigmetrics Test of Time Award in 2011, and has over 3600 citations. In a susequent work, Sanjay and his colleagues developed a prototype overlay multicast system which was used to broadcast several tens of events including ACM Sigcomm and ACM SOSP (the experience described in a Usenix ATC 2004 paper).