May 2022 - talks - lists.cs.princeton.edu

CS Colloquium Series: week of April 11-15
by Emily C. Lawrence 19 Sep '22

19 Sep '22

Here is the full list of CS Colloquium talks for next week. All talks will be recorded. ~~~~~ CS Colloquium Speaker Speaker: Parastoo Abtahi, Stanford University Date: Monday, April 11, 2022 Time: 12:30pm EST Location: CS 105 Hosts: Andrés Monroy-Hernández & Adam Finkelstein Event page: [ https://www.cs.princeton.edu/events/26186 | https://www.cs.princeton.edu/events/26186 ] This talk will be live-streamed at [ https://mediacentrallive.princeton.edu/ | https://mediacentrallive.princeton.edu/ ] Title: From Haptic Illusions in Virtual Reality to Beyond-Real Interactions Abstract: Advances in audiovisual rendering have led to the commercialization of virtual reality (VR) hardware; however, haptic technology has not kept up with these advances. While haptic devices aim to bridge this gap by simulating the sensation of touch, there are many hardware limitations that make realistic touch interactions in VR challenging. In my research, I explore how by understanding human perception, we can design VR interactions that not only overcome the current limitations of VR hardware, but also extend our abilities beyond what is possible in the real world. In this talk, I will present my work on redirection illusions that leverage the limits of human perception to improve the perceived performance of encountered-type haptic devices, such as improving the position accuracy of drones, the speed of tabletop robots, and the resolution of shape displays when used for haptics in VR. I will then present a framework I have developed through the lens of sensorimotor control theory to argue for the exploration and evaluation of VR interactions that go beyond mimicking reality. Bio: Parastoo Abtahi is a final year computer science PhD candidate and a Gerald J. Lieberman fellow at Stanford University, where she is co-advised by Prof. James Landay and Prof. Sean Follmer. Her research area is human-computer interaction (HCI) and she works broadly on virtual reality interactions and spatial computing. Her research has been published at top HCI venues, such as the ACM Conference on Human Factors in Computing Systems (CHI) and the ACM User Interface Software and Technology Symposium (UIST), and has received two honorable mention paper awards. Her work has been supported by the Stanford Institute for Human-Centered Artificial Intelligence, the Hasso Plattner Design Thinking Research Program, and the VMware fellowship. Prior to Stanford, Parastoo received her bachelor’s degree in Electrical and Computer Engineering from the University of Toronto as part of the Engineering Science program. ~~~~~ CS Colloquium Speaker Speaker: Ferran Alet, Massachusetts Institute of Technology Date: Tuesday, April 12, 2022 Time: 12:30pm EST Location: CS 105 Host: Olga Russakovsky Event page: https://www.cs.princeton.edu/events/26187 This talk will be live-streamed at https://mediacentrallive.princeton.edu/ Title: A flexible framework for machine learning Abstract: In this last decade, we have seen a lot of progress in AI and machine learning using a single recipe: given a task, we train a single neural network to map inputs to outputs. In this talk, I will show that this one-neural-network-per-task framework can be extended to improve generalization. First, I will describe modular meta-learning, which achieves language-like generalization by training a set of composable neural modules. By having multiple neural networks per task, and multiple tasks per neural network, we are able to reuse information and achieve bigger data and computational efficiency. In the second part of my talk, I will describe tailoring, a general way of encoding inductive biases in neural networks by optimizing unsupervised objectives inside the prediction function, essentially having one neural network per input. Finally, I will describe my vision for creating a flexible ML framework that will enable training reinforcement learning policies within minutes rather than days, solving complex search and discovery problems, and improving our understanding of generalization in deep learning. Bio: Ferran Alet is a PhD candidate at MIT CSAIL advised by Leslie Kaelbling, Tomas Lozano-Perez, and Joshua Tenenbaum. His research is on machine learning and leverages techniques from meta-learning, learning to search, program synthesis, and insights from mathematics and the physical sciences. During his PhD, he created the MIT Embodied Intelligence Seminar, mentored 17 students, and won the MIT Outstanding Mentor award 2021. Ferran studied mathematics and physics in Barcelona thanks to CFIS, a program for doing two degrees, where he was the valedictorian of his promotion. In college, he participated in the ACM-ICPC programming contest, being the most decorated in the history of his regional phase (South Western Europe). In grad school, he earned a “La Caixa” fellowship and was responsible for the high-level planner of the MIT-Princeton team for the Amazon Robotics Challenge, which won the stowing task in 2017. You can find more information and papers at www.alet-et.al ~~~~~ CS Colloquium Speaker Speaker: Benedikt Bünz, Stanford University Date: Thursday, April 14, 2022 Time: 12:30pm EST Location: CS 105 Hosts: Jonathan Mayer & Arvind Narayanan Event page: https://www.cs.princeton.edu/events/26188 This talk will be live-streamed at https://mediacentrallive.princeton.edu/ Title: Improving the privacy, scalability, and ecological impact of blockchains Abstract: Blockchains are an exciting area of research that touches on many areas of Computer Science and beyond. This technology has the potential to enable a fast, cheap, and private financial system based on distributed consensus and cryptography, instead of trusted parties. Despite this potential, the reality still shows severe limitations of blockchains: (i) transactions can cost hundreds of dollars and take minutes to confirm, (ii) some blockchains offer little privacy, and (iii) proof-of-work consensus consumes too much energy. In this talk, I will discuss powerful techniques that follow a prover paradigm and can mitigate these limitations. The first technique, called Bulletproofs, is a general-purpose zero-knowledge proof system that is specifically designed to enable confidential blockchain transactions. Bulletproofs requires minimal trust assumptions and gives the shortest zero-knowledge proofs without trusted setup. The system is widely deployed and powers tens of thousands of private blockchain transactions per day. The second technique, called inner pairing products, is a way to aggregate many zero-knowledge proofs into a single short proof. This can significantly reduce on-chain data, leading to a significant increase in transactions per second that the chain can process. The third technique is a new concept called a verifiable delay function (VDF) that is vital for permission-less and eco-friendly consensus. VDFs are already deployed in Filecoin and Chia, and are planned for Ethereum 2.0, the upcoming upgrade to Ethereum. Bio: Benedikt Bünz is a PhD candidate at Stanford University, a member of Dan Boneh’s applied cryptography lab, and a recipient of the Microsoft Research Fellowship at the Simons Institute. His work on the science of Blockchains uses tools from applied cryptography, distributed systems, and algorithmic game theory. His research focuses on building new proof protocols for improving the privacy, scalability, and ecological impact of blockchains. Several of his research results have had a significant industry impact. His work on Bulletproofs, secures tens of thousands of private transactions on Blockchains like Monero or Signal’s Mobilecoin. His seminal work on Verifiable Delay Functions (VDFs) sparked the VDF Alliance, a multi-million dollar initiative composed of academic, non-profit, and corporate collaborators.

1 11

ECE/COS Seminar: Adrian Perrig, ETH Zurich, Thursday, May 26 at 2:00pm
by Emily C. Lawrence 26 May '22

26 May '22

ECE Korhammer Seminar Series/COS Colloquium Series Speaker: Adrian Perrig, ETH Zurich Title: Experiencing a new Internet Architecture Date: Thursday, May 26, 2022 Time: 2:00 pm Location: B205 Engineering Quadrangle Hosts: Prateek Mittal & Jen Rexford Event page: [ https://ece.princeton.edu/events/experiencing-new-internet-architecture | https://ece.princeton.edu/events/experiencing-new-internet-architecture ] Virtual option available via Zoom: [ https://princeton.zoom.us/j/95987490029?pwd=cDRIMkk1Q2FadVpnbk9BSWpBQUg3dz09 | https://princeton.zoom.us/j/95987490029?pwd=cDRIMkk1Q2FadVpnbk9BSWpBQUg3dz09 ] Abstract: Imagining a new Internet architecture enables us to explore new networking concepts without the constraints imposed by the current infrastructure. What are the benefits of a multi-path inter-domain routing protocol that finds dozens of paths? What about a data plane without inter-domain forwarding tables on border routers? What secure systems can we build if a router can derive a symmetric key for any host on the Internet within nanoseconds? In this presentation, we invite you to join us on our 12-year long expedition of creating the SCION next-generation secure Internet architecture. SCION has already been deployed at several ISPs and domains, and has been in production use since 2017. On our journey, we have found that path-aware networking and multipath communication not only provide security benefits, but also enable higher efficiency for communication, increase network capacity, and can even provide opportunities for reducing the carbon footprint of communication. Bio: Adrian Perrig is a Professor at the Department of Computer Science at ETH Zürich, Switzerland, where he leads the network security group. He is also a Distinguished Fellow at CyLab, and an Adjunct Professor of Electrical and Computer Engineering at Carnegie Mellon University. From 2002 to 2012, he was a Professor of Electrical and Computer Engineering, Engineering and Public Policy, and Computer Science (courtesy) at Carnegie Mellon University. From 2007 to 2012, he served as the technical director for Carnegie Mellon's Cybersecurity Laboratory (CyLab). He earned his MS and PhD degrees in Computer Science from Carnegie Mellon University, and spent three years during his PhD at the University of California at Berkeley. He received his BSc degree in Computer Engineering from EPFL. He is a recipient of the ACM SIGSAC Outstanding Innovation Award. Adrian is an ACM and IEEE Fellow. Adrian's research revolves around building secure systems -- in particular his group is working on the SCION secure Internet architecture. This seminar is supported with funds from the Korhammer Lecture Series. Sponsored by the departments of Electrical & Computer Engineering and Computer Science.

1 1

Andy Jones will present his Pre FPO "Probabilistic models for structured biomedical data" on May 26, 2022 at 10am via Zoom
by Nicki Mahler 19 May '22

19 May '22

Andy Jones will present his Pre FPO "Probabilistic models for structured biomedical data" on May 26, 2022 at 10am via Zoom. Zoom link: https://princeton.zoom.us/j/91964333824 Title Probabilistic models for structured biomedical data Abstract It is commonly assumed in biomedical data analysis that the samples are independent and identically distributed. However, this assumption is nearly always violated in practice. Data are often composed of subgroups of samples (e.g., patients, cells, or tissue types) and usually contain substantial correlation structure among the samples and features. In this body of work, we propose a series of data modeling and analysis frameworks to account for these highly-structured experimental data. First, we propose two methods for performing dimension reduction and exploratory data analysis in a case-control setting. Second, we propose a spatial alignment method that registers spatial genomics slices into a common coordinate frame. Finally, in ongoing work, we propose a family of methods for optimally designing spatial genomics experiments. In each of these settings, we demonstrate the benefits of exploiting known experimental structure for data analysis and downstream findings. References Published or under review: Alignment of spatial genomics and histology data using deep Gaussian processes. Andrew Jones, F. William Townes, Didong Li, Barbara E. Engelhardt (2022). Multi-group Gaussian Processes. Didong Li, Andrew Jones, Sudipto Banerjee, Barbara E. Engelhardt (2021). Contrastive latent variable modeling with application to case-control sequencing experiments. Andrew Jones, F. William Townes, Didong Li, Barbara E. Engelhardt. The Annals of Applied Statistics (to appear). Nested policy reinforcement learning. Aishwarya Mandyam, Andrew Jones, Krzysztof Laudanski, Barbara E. Engelhardt (2021). Efficient Bayesian Inverse Reinforcement Learning via Conditional Kernel Density Estimation. Aishwarya Mandyam, Didong Li, Diana Cai, Andrew Jones, Barbara E. Engelhardt Probabilistic Contrastive Principal Component Analysis. Didong Li*, Andrew Jones*, Barbara E. Engelhardt (2020). In preparation: Optimizing experimental design for spatial transcriptomics studies. Andrew Jones, Barbara E. Engelhardt. Spatially-aware dimension reduction via semi-supervised latent variable models. Lauren Okamoto, Andrew Jones, Archit Verma, Barbara E. Engelhardt. A Poisson reduced rank regression model for association mapping in sequencing data. Tiana Fitzgerald, Andrew Jones, Barbara E. Engelhardt. Linking histology and molecular state across human tissues. Andrew Jones, Gregory W. Gundersen, Barbara E. Engelhardt.

1 0

David Liu will present his Pre-FPO "A Serverless Architecture for Application-level Orchestration" on Tuesday, May 17, 2022 at 2:00 PM via Zoom.
by Louis Riehl 13 May '22

13 May '22

David Liu will present his Pre-FPO "A Serverless Architecture for Application-level Orchestration" on Tuesday, May 17, 2022 at 2:00 PM via Zoom. Zoom link: https://princeton.zoom.us/j/92893120748 Committee Members: Amit Levy (Examiner, advisor), Mike Freedman (Examiner, advisor), Jennifer Rexford(examiner), Wyatt Lloyd (reader), Ravi Netravali (reader). All are welcome to attend. Title: A Serverless Architecture for Application-level Orchestration Abstract: Serverless workflow orchestrators make it easier for developers to build complex serverless applications by driving the execution of a workflow from a central service. Making orchestrators fault-tolerant and scalable requires engineering effort as well as large up-front resource costs to deploy. As a result, developers typically rely on multi-tenant services run by cloud providers that can amortize these costs across many users. Unfortunately, this centralization comes at the expense of application flexibility. Orchestrators support a fixed set of application patterns, and trade off performance, scalability, fault tolerance, and resource efficiency using a policy chosen by the cloud provider. They support many applications well, but cannot simultaneously satisfy the needs of all applications. We argue that centralized serverless orchestrators are actually unnecessary. Instead, a decentralized orchestrator, deployed as a library included by the serverless application, can support the same application patterns and fault tolerance as existing orchestrators using only the basic building blocks of serverless platforms---a scalable ``Function'' executor (FaaS) and a consistent scalable datastore. This enables different applications to use or customize different libraries to implement different patterns or make application-specific trade-offs. To demonstrate this, we describe the design and implementation of Unum, a decentralized serverless workflow system. Unum compiles workflows written in existing high-level workflow languages into a partitioned orchestration library that executes orchestration logic in-situ with user functions. Unum provides the same strong correctness guarantees under faults as a centralized orchestrator, performs similarly, scales better, and costs less than existing centralized workflow systems. Our implementation of Unum supports a superset of applications the AWS Step Functions orchestrator supports, runs on both AWS Lambda and Google Cloud Functions, and runs up to 2x faster and up to 9x cheaper than native Step Functions on a representative set of applications.

1 0

Dongsheng Yang will present his General Exam on Monday, May 9, 2022 at 2pm via Zoom.
by Nicki Mahler 09 May '22

09 May '22

Dongsheng Yang will present his General Exam on Monday, May 9, 2022 at 2pm via Zoom. Zoom link: princeton.zoom.us/j/6605501024 The members of his committee are as follows: Kai Li (adviser), Wyatt Lloyd, Ravi Netravali Abstract: Recent work shows the effectiveness of Machine Learning (ML) to reduce cache miss ratios by making better eviction decisions than heuristics. However, state-of-the-art ML caches require many predictions to make an eviction decision, making them not practical for high-throughput caching systems. We introduce Machine learning At the Tail (MAT), a framework to build efficient ML-based caching systems by integrating with a heuristic algorithm. MAT treats the heuristic algorithm as a “filter” to select high- quality samples to train an ML model and likely candidate objects for evictions. We evaluate MAT on 8 production workloads, spanning storage, in-memory caching, and CDNs. The experiments show that MAT can reduce the number of ML predictions of the state-of-the-art ML-based cache from 63 to 2, reducing the software overhead by 31 times, while achieving similar miss ratios. Link to Reading List: [ https://docs.google.com/document/d/1WeBtRlmeY9Oulz6nhGeFoAYKPb4IC6tzkWmKX0g… | https://docs.google.com/document/d/1WeBtRlmeY9Oulz6nhGeFoAYKPb4IC6tzkWmKX0g… ] Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 1

Sophia Yoo (ECE) will present her General Exam "SmartCookie: Distributed In-Network SYN Flooding Mitigation" on Tuesday, May 10, 2022 at 9am in CS 302.
by Nicki Mahler 09 May '22

09 May '22

Sophia Yoo (ECE) will present her General Exam "SmartCookie: Distributed In-Network SYN Flooding Mitigation" on Tuesday, May 10, 2022 at 9am in CS 302. Title: SmartCookie: Distributed In-Network SYN Flooding Mitigation Date: Tuesday, May 10, 2022 Time: 9 am - 10 am Location: CS 302 Committee: Jen Rexford, Prateek Mittal, Ravi Netravali Abstract: SYN-flooding attacks, a pervasive class of DDoS attacks, target the TCP three-way handshake to exhaust server resources, leading to poor performance for legitimate clients. In this work, we present SmartCookie, a distributed defense system for service providers to protect their servers from large-scale SYN flooding attacks. SmartCookie pushes the front line of defense to the network edge, placing an in-network SYN cookie proxy upstream of the servers and closer to the clients. Our design capitalizes on the distinctive capabilities of high-speed programmable switches by offloading some defense responsibilities to lightweight modules running within standard Linux kernels. Unlike traditional SYN cookie proxy defenses located at the server, SmartCookie is a split proxy with a novel division of labor between a high-speed, in-network switch and a lightweight server-side component. Our design protects downstream network bandwidth (by blocking attack traffic early in the path), scales to protect many servers at a time, and gracefully handles asymmetric routing. We implement SmartCookie with P4 (on Intel Tofino switches) and eBPF (on Linux servers), showing two orders of magnitude improvement over the standard Linux kernel's defense running on an 8-core CPU, and blocking attack rates of 136.9 million packets per second with no packet loss.

1 1

Olga Solodova will present her General Exam TODAY, Monday, May 9, 2022 at 11:00 AM in CS 402 and via zoom.
by Louis Riehl 09 May '22

09 May '22

Olga Solodova will present her General Exam on Monday, May 9, 2022 at 11:00 AM in CS 402 and via zoom. Zoom link: https://princeton.zoom.us/j/6387833692?pwd=a25TVlhkWlFYUE9uQ2dyVUI4eGpDdz09 Committee Members: Ryan Adams (advisor), Adji Bousso Dieng, Szymon Rusinkiewicz Abstract: Two projects will be presented in this work. The first uses deep reinforcement learning (RL) to generate optimized assembly code for arithmetic circuits. The second aims to add to methods for graph-based learning. Arithmetic circuits are ubiquitous in machine learning and scientific computing; computation of polynomial functions, matrix multiplication, and solutions to differential equations are all arithmetic circuits. These algorithms are so central and executed so often that it is imperative that their implementation is optimal (in terms of metrics such as execution speed or energy consumption). In practice, hundreds of human hours are invested in directly writing optimized machine code for such applications. Unfortunately, machine code implementations of these algorithms are difficult to reason about, especially given increasingly sophisticated computer architectures which have complicated performance characteristics. In this work, Monte Carlo tree search together with a policy and value network is used to efficiently search for optimal solutions in the space of assembly programs which compute some arithmetic circuit. Graph neural networks (GNNs) are neural networks designed to work with graph-structured data such as the arithmetic circuits described above. Specifically, GNNs generate node embeddings by repeatedly aggregating information from the neighborhood around each node in a graph, and use these embeddings to calculate a prediction. The aggregation and prediction functions consist of neural networks with parameters which are learned from data. At a high level, there are two approaches to learning in GNNs. The first is to have the aggregation function converge to a fixed point in the embedding space of the nodes, and employ the Almeida-Pineda algorithm to take gradients at this fixed point. The second is to unroll the aggregation function for a set number of time steps and use back-propagation through time (BBPT). In existing GNNs, employing the first approach requires constraining the aggregation function generating node embeddings to be a contraction map in order to arrive at a fixed point. This restriction limits the propagation of information across long distances in a graph, which diminishes the ability to learn long-range dependencies. However, having node embeddings converge to a fixed point removes the need to 'unroll' all the iterations of the fixed point iteration, which is particularly useful in memory-constrained settings when the graph or number of fixed point iterations is large. The second approach using BBPT suffers from different problems, namely that 'over-smoothing' in the node embeddings is difficult to mitigate in deep GNNs (i.e. GNNs with many iterations of the aggregation function). This work presents energy-based GNNs, which use the minimization of an energy function as the aggregation function, which by definition arrives at a fixed point in the space of node embeddings. This allows employing the Almeida-Pineda algorithm to take gradients, and may additionally enable more effective node embeddings to be learned with deep GNNs. Reading List: https://docs.google.com/document/d/1qnjynlegkn29HEzg6mS6bO-aAodUxNxdA5S4bqUB AV0/edit Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Nick Richardson will present his General Exam "Fiber Sampling: Differentiable Monte Carlo for Geometric Objectives" on May 11, 2022 at 2pm via Zoom and CS 402
by Nicki Mahler 06 May '22

06 May '22

Nick Richardson will present his General Exam "Fiber Sampling: Differentiable Monte Carlo for Geometric Objectives" on May 11, 2022 at 2pm via Zoom and CS 402 Zoom link: [ https://us02web.zoom.us/j/84741197027?pwd=STBHTFd0WnpBeVIrTjF0d2dhbkZSUT09 | https://us02web.zoom.us/j/84741197027?pwd=STBHTFd0WnpBeVIrTjF0d2dhbkZSUT09 ] The members of his committee are as follows: Ryan Adams (adviser), Adji Bousso Dieng, and Szymon Rusinkiewicz Abstract: Integrals with parametric, discontinuous integrands are ubiquitous, and arise via the discrete latent structure resident in applications like shape optimization, phys- ical simulation, and graphics. Treating these tasks as inverse problems with an associated generative process requires optimizing model parameters which inter- act via these discontinuities. The advent of powerful automatic differentiation libraries teases the use of generic gradient-based optimization to perform inference in these models, but this requires computing derivatives of integrals with paramet- ric discontinuities. A systematic methodology to wield conventional automatic differentiation frameworks in these discrete contexts is wanting. In this work, we engage with the difficulty of computing these derivatives by noting a dual between the geometric/analytic properties of these functions and the probabilistic view of integration, namely Monte Carlo estimation. We introduce a differentiable variant of the simple Monte Carlo estimator which facilitates the general computation of these challenging derivatives, utilizing conventional automatic differentiation frameworks. We justify the estimator analytically, and demonstrate the generality of the method as applied to differentiable rendering, implicit shape optimization, PDE solving, and volumetric reconstruction. Reading List: https://drive.google.com/drive/u/1/folders/1VT6kaFKBOnprydP0a_ACaH45IFKcz-ad Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

UPDATE: Yunyun Wang will present her General Exam "Controllable Speech Representation Learning via Voice Conversion" Wednesday, May 11, 2022 at 1:30 pm via Zoom
by Louis Riehl 06 May '22

06 May '22

Yunyun Wang will present her General Exam "Controllable Speech Representation Learning via Voice Conversion" on Wednesday, May 11, 2022 at 1:30 pm via Zoom Zoom Link: https://princeton.zoom.us/j/9195627075?pwd=a1N5a2VzYy92cEFMVjFTZUZnOGpiUT09 Committee Members: Adam Finkelstein (advisor), Karthik Narasimhan, Olga Russakovsky Abstract: Speech representation learning transforms speech into features that are suitable for downstream tasks, such as speech recognition, phoneme classification or speaker identification. For such recognition tasks, a representation can be lossy (non-invertible), which is typical of BERT-like, self-supervised models. However, when used for synthesis tasks, we find these lossy representations prove to be insufficient to plausibly reconstruct the input signal. This paper introduces a method for invertible and controllable speech representation learning based on disentanglement. The representation can be decoded into a signal perceptually identical to the original. Moreover, its disentangled components (speech content, pitch, speaker identity and energy) can be controlled independently to alter the synthesis result. Our model builds upon a zero-shot model called AutoVC-F0 trained for voice conversion – where the goal is to modify an audio recording containing the voice of one speaker so that the identity sounds like that of another speaker without altering the speech content. Building on AutoVC-F0, we introduce alteration invariant content loss (AIC loss) as well as adversarial training (GAN). Through objective measures and subjective tests, we show that our formulation offers significant improvement in voice conversion sound quality as well as more precise control over the disentangled features. Reading List: https://docs.google.com/document/d/1YTGrKke9HYHa7bxtNiyTyHpArrwSEXulPmrb_Dw… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Aninda Manocha will present her Pre FPO "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" on May 12, 2022 at 2pm via Zoom
by Nicki Mahler 06 May '22

06 May '22

Aninda Manocha will present her Pre FPO "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" on May 12, 2022 at 2pm via Zoom. Zoom link: [ https://princeton.zoom.us/j/96476041500 | https://princeton.zoom.us/j/96476041500 ] Committee: Margaret Martonosi (adviser), Readers: Kai Li and Luca Carloni, Examiners: David August and David Wentzlaff All are welcome to attend. Title: "Optimizing Data Supply and Memory Management for Graph Applications in Post-Moore Hardware-Software Systems" Abstract: Graph structures naturally and efficiently capture relationships between entities, such as individuals in a social network, pages in the World Wide Web, and amino acids in protein molecules. Therefore, graph processing algorithms are commonly used in many important big-data applications. However, modern accelerator-oriented, heterogeneous system designs are inefficient for graph applications because they are characterized by irregular memory access patterns that confound sophisticated instruction scheduling and caching mechanisms as well as traditional prefetching and speculation techniques. As a result, these applications are severely bottlenecked by several long-latency memory accesses to off-chip DRAM and address translation overheads. My thesis aims to tackle these bottlenecks with three main thrusts spanning the hardware-software stack. First, I propose a hardware-software co-design composed of specialized access tracking buffers and precise compiler-driven program slicing techniques. This design enables significant memory-level parallelism and performance gains by issuing and overlapping multiple long-latency memory accesses to amortize their cost. This approach outperforms out-of-order cores, do-all parallelism, prefetching, and prior decoupling approaches, achieving a 2.87x speedup and 8.61x gain in energy efficiency across a range of graph applications. Second, I present domain-specific cache management techniques that leverage tailored fetch, insertion, and replacement policies to alleviate memory bandwidth pressure and minimize off-chip DRAM accesses. This work yields up to 1.79x (geomean 1.3x) speedup over state-of-the-art cache management policies. Lastly, I describe application-specific operating system (OS) techniques to efficiently manage page sizes and mitigate address translation overheads. These mechanisms offer 77.3-96.3% the performance of ideal huge page usage, while requiring only 0.58-2.92% of the huge page resources that modern operating systems, namely Linux, require. Overall, my work demonstrates the need for architects to revisit and equip simple hardware with specialized data supply technology to accelerate memory-bound applications when Dennard scaling no longer helps performance and energy savings are critical. Furthermore, as the end of Moore’s Law has pushed computer architects to employ hardware accelerators in order to reach aggressive power and performance targets for compute-bound workloads, domain-specific cache hierarchy approaches and OS techniques for memory-bound workloads are becoming increasingly necessary. Together, the three proposals in my dissertation work attack the memory bottlenecks in graph applications from different perspectives of the hardware-software stack to enable optimal performance.

1 0