A5058729992- Software-Defined Networks and 5G
- Optical Network Technologies
- Advanced Optical Network Technologies
- Interconnection Networks and Systems
- Neural Networks and Reservoir Computing
- Photonic and Optical Devices
- Cloud Computing and Resource Management
- Advanced Photonic Communication Systems
- Network Traffic and Congestion Control
- Advanced Memory and Neural Computing
- Privacy-Preserving Technologies in Data
- Parallel Computing and Optimization Techniques
- Adversarial Robustness in Machine Learning
- Caching and Content Delivery
- IoT and Edge/Fog Computing
- Distributed and Parallel Computing Systems
- Network Security and Intrusion Detection
- Cryptography and Data Security
- Complex Network Analysis Techniques
- Semiconductor Lasers and Optical Devices
- Stochastic Gradient Optimization Techniques
- Distributed systems and fault tolerance
- Advanced Neural Network Applications
- Advanced Data Storage Technologies
- Graph Theory and Algorithms
Massachusetts Institute of Technology
2019-2024
IIT@MIT
2020-2024
Vassar College
2022-2023
Moscow Institute of Thermal Technology
2020-2023
Cambridge Electronics (United States)
2022
Columbia University
2019-2022
Lawrence Berkeley National Laboratory
2019-2022
Fujitsu (Japan)
2022
Institut Polytechnique de Paris
2022
Intel (United States)
2022
Advanced machine learning models are currently impossible to run on edge devices such as smart sensors and unmanned aerial vehicles owing constraints power, processing, memory. We introduce an approach inference based delocalized analog processing across networks. In this approach, named Netcast, cloud-based "smart transceivers" stream weight data devices, enabling ultraefficient photonic inference. demonstrate image recognition at ultralow optical energy of 40 attojoules per multiply (<1...
To keep up with the continuous growth in demand, cloud providers spend millions of dollars augmenting capacity their wide-area backbones and devote significant effort to efficiently utilizing WAN capacity. A key challenge is striking a good balance between network utilization availability, as these are inherently at odds; highly utilized might not be able withstand unexpected traffic shifts resulting from link/node failures. We advocate novel approach this that draws inspiration financial...
Fiber optic cables connecting data centers are an expensive but important resource for large organizations. Their importance has driven a conservative deployment approach, with redundancy and reliability baked in at multiple layers. In this work, we take more aggressive approach argue adapting the capacity of fiber links based on their signal-to-noise ratio (SNR). We investigate idea by analyzing SNR over 8,000 optical backbone period three years. show that 64% 100 Gbps IP can be augmented...
This paper proposes optical network interconnects as a key enabler for building high-bandwidth ML training clusters with strong scaling properties. Our design, called SiP-ML, accelerates the time of popular DNN models using silicon photonics links capable providing multiple terabits-per-second bandwidth per GPU. SiP-ML partitions job across GPUs hybrid data and model parallelism while ensuring communication pattern can be supported efficiently on interconnect. We develop task partitioning...
This paper describes Trio, a programmable chipset used in Juniper Networks' MX-series routers and switches. Trio's architecture is based on multi-threaded packet processing engine hierarchy of high-capacity memory systems, making it fundamentally different from pipeline-based architectures. Trio gracefully handles non-homogeneous rates for wide range networking use cases protocols, an ideal platform emerging in-network applications. We begin by describing the chipset's fundamental building...
Emerging data center architectures are becoming reconfigurable. While prior work has shown the practical benefits of reconfigurable topologies, underlying algorithmic complexity is not yet well understood. In particular, most topologies hybrid, where parts network (consisting optical or wireless devices) while other static electrical switches). Current proposals enforce a routing policy that routes flows on either part "exclusively" by labeling as mice elephant. We show such artificial...
This paper presents a systematic approach to identify and quantify the types of structures featured by packet traces in communication networks. Our leverages an information-theoretic methodology, based on iterative randomization compression trace, which allows us systematically remove measure dimensions structure trace. In particular, we introduce notion trace complexity approximates entropy rate Considering several real-world traces, show that can provide unique insights into...
Today's network devices share buffer across queues to avoid drops during transient congestion and absorb bursts. As the buffer-per-bandwidth-unit in datacenter decreases, need for optimal utilization becomes more pressing. Typical use a hierarchical packet admission control scheme: First, Buffer Management (BM) scheme decides maximum length per queue at device level then an Active Queue (AQM) which packets will be admitted level. Unfortunately, lack of cooperation between two schemes leads...
Programmable switches based on the Protocol Independent Switch Architecture (PISA) have greatly enhanced flexibility of today's networks by allowing new packet protocols to be deployed without any hardware changes. They also been instrumental in enabling a computing paradigm which parts an application's logic run within network core (in-network computing).
As bytes-per-FLOP ratios continue to decline, communication is becoming a bottleneck for performance scaling. This paper describes bandwidth steering in HPC using emerging reconfigurable silicon photonic switches. We demonstrate that placing photonics the lower layers of hierarchical topology efficiently changes connectivity and consequently allows operators recover from system fragmentation otherwise hard mitigate common task placement strategies. Bandwidth enables efficient utilization...
We review the motivation, goals, and achievements of Photonic Integrated Networked Energy efficient datacenter (PINE) project, which is part Advanced Research Projects Agency–Energy (ARPA-E) ENergy-efficient Light-wave Technology Enabling Networks that Enhance Dataprocessing (ENLITENED) program. The PINE program leverages unique features photonic technologies to enable alternative mega-datacenters high-performance computing (HPC) system architectures deliver more substantial energy...
The bandwidth and latency requirements of modern datacenter applications have led researchers to propose various topology designs using static, dynamic demand-oblivious (rotor), and/or demand-aware switches. However, given the diverse nature traffic, there is little consensus about how these would fare against each other. In this work, we analyze throughput existing under different traffic patterns study their unique advantages potential costs in terms ''tax''. To overcome identified...
Designing efficient interconnects to support high-bandwidth and low-latency communication is critical toward realizing high performance computing (HPC) data center (DC) systems in the exascale era. At extreme scales, providing requisite bandwidth through overprovisioning becomes impractical. These challenges have motivated studies exploring reconfigurable network architectures that can adapt traffic patterns at runtime using optical circuit switching. Despite plethora of proposed...
The massive growth of machine learning-based applications and the end Moore's law have created a pressing need to redesign computing platforms. We propose Lightning, first reconfigurable photonic-electronic smartNIC serve real-time deep neural network inference requests. Lightning uses fast datapath feed traffic from NIC into photonic domain without creating digital packet processing data movement bottlenecks. To do so, leverages novel count-action abstraction that keeps track required...
Fiber cut events reduce the capacity of wide-area networks (WANs) by several Tbps. In this paper, we revive lost reconfiguring wavelengths from fibers into healthy fibers. We highlight two challenges that made prior solutions impractical and propose a system called Arrow to address them. First, our measurements show contrary common belief, in most cases, is only partially restorable. This poses cross-layer challenge Traffic Engineering (TE) perspective has not been considered before: "Which...
This paper argues that fair-sharing, the holy grail of congestion control algorithms for decades, is not necessarily a desirable property in Machine Learning (ML) training clusters. We demonstrate specific combination jobs, introducing unfairness improves time all competing jobs. call this jobs compatible and define compatibility criterion using novel geometric abstraction. Our abstraction rolls around circle rotates communication phases to identify fully Using abstraction, we up 1.3×...
This paper presents a systematic approach to identify and quantify the types of structures featured by packet traces in communication networks. Our leverages an information-theoretic methodology, based on iterative randomization compression trace, which allows us systematically remove measure dimensions structure trace. In particular, we introduce notion \emphtrace complexity approximates entropy rate Considering several real-world traces, show that trace can provide unique insights into...
Conventional buffer sizing techniques consider an output port with multiple queues in isolation and provide guidelines for the size of queue. In practice, however, switches consist several ports that share a buffering chip. Hence, chip manufacturers, such as Broadcom, are left to devise set proprietary resource sharing algorithms allocate buffers across ports. This algorithm dynamically adjusts directly impacts packet loss latency individual queues. We show problem allocating ports, although...
We propose TopoOpt, a novel direct-connect fabric for deep neural network (DNN) training workloads. TopoOpt co-optimizes the distributed process across three dimensions: computation, communication, and topology. demonstrate mutability of AllReduce traffic, leverage this property to construct efficient topologies DNN jobs. then uses an alternating optimization technique group theory-inspired algorithm called TotientPerms find best topology routing plan, together with parallelization strategy....
We present MLTCP, a technique to augment today's congestion control algorithms accelerate DNN training jobs in shared GPU clusters. MLTCP enables the communication phases of that compete for network bandwidth interleave with each other, thereby utilizing efficiently. At heart lies very simple principle based on key conceptual insight: flows should scale their window size number bytes sent at iteration. show integrating this into protocols is straightforward: by adding 30-60 lines code Reno,...