Abstract

Overview of the weekly meeting of computing lunch. We will determine presenters in this semester.

Abstract

The recent success of AI / Machine Learning is largely due to the availability of Big data and the infrastructure to scale computation. As a result, the integration of AI / Machine Learning and Big data management techniques becomes inevitable. The Data Intelligence lab performs research on this integration in both directions. First, we explore Machine Learning techniques that can improve Big data management. Second, we explore Big data management techniques that are needed throughout a Machine Learning lifecycle. In this computing lunch talk, we will cover existing research in these directions and opportunities for new research.

Abstract

As the complexity of modern networks increases, virtualization techniques, such as software-defined networking (SDN) and network function virtualization (NFV), get highlighted to achieve various network management and operating requirements. However, those virtualization techniques (specifically, NFV) have a critical issue that the performance of virtualized network functions (VNFs) is easily affected by diverse environmental factors (e.g., various workloads, resource contentions among VNFs), so resulting in unexpected performance degradations - performance uncertainty. Unfortunately, existing approaches mostly provide limited information about a single VNF or the underlying infrastructure (e.g., Xen, KVM), which is deficient in reasoning why the performance uncertainties occur. For such reasons, we first deeply investigate the behaviors of multiple VNFs along service chains in NFV environments, and define a set of critical performance features for each layer in the NFV hierarchical stack. Based on our investigations and findings, we introduce an automated analysis system, Probius, providing the comprehensive view of VNFs and their service chains on the basis of NFV architectural characteristics. Probius collects most possible NFV performance related features efficiently, analyzes the behaviors of NFV, and finally detects abnormal behaviors of NFV - possible reasons of performance uncertainties. To show the effectiveness of Probius, we have deployed 7 open-source VNFs and found 5 interesting performance issues caused by environmental factors.

Abstract

In this talk, I will present Atom, an anonymous messaging system that protects against traffic-analysis attacks. Unlike many prior systems, each Atom server touches only a small fraction of the total messages routed through the network. As a result, the system?™s capacity scales near-linearly with the number of servers. At the same time, each Atom user benefits from ?œbest possible??anonymity: a user is anonymous among all honest users of the system, even against an active adversary who monitors the entire network, a portion of the system?™s servers, and any number of malicious users. The architectural ideas behind Atom have been known in theory, but putting them into practice requires new techniques for (1) avoiding heavy general purpose multi-party computation protocols, (2) defeating active attacks by malicious servers at minimal performance cost, and (3) handling server failure and churn. Atom is most suitable for sending a large number of short messages, as in a microblogging application. We show that, on a heterogeneous network of 1,024 servers, Atom can transit a million Tweet-length messages in 28 minutes. This is over 23× faster than prior systems with similar privacy guarantees.

Abstract

The problem of detecting whether a test sample is from in-distribution (i.e., training distribution by a classifier) or out-of-distribution sufficiently different from it arises in many real-world machine learning applications. However, the state-of-art deep neural networks are known to be highly overconfident in their predictions, i.e., do not distinguish in- and out-of-distributions. Recently, to handle this issue, several threshold-based detectors have been proposed given pre-trained neural classifiers. However, the performance of prior works highly depends on how to train the classifiers since they only focus on improving inference procedures. In this paper, we develop a novel training method for classifiers so that such inference algorithms can work better. In particular, we suggest two additional terms added to the original loss (e.g., cross entropy). The first one forces samples from out-of-distribution less confident by the classifier and the second one is for (implicitly) generating most effective training samples for the first one. In essence, our method jointly trains both classification and generative neural networks for out-of-distribution. We demonstrate its effectiveness using deep convolutional neural networks on various popular image datasets.

Abstract

Many computational problems can be conveniently expressed in terms of some common logical frameworks, and this can often lead to effective solution approaches because we can concentrate the development of efficient algorithms for a few general-purpose problems. Boolean satisfiability (SAT) solving is the best known example of this paradigm, but generalizations such as satisfiability modulo theories (SMT) also have wide application. My research extends this paradigm to what are called model counting problems (#SAT), in which the goal is not just to find a single solution to a set of constraints but to count, perhaps approximately, the total number of solutions.

Abstract

In this talk, I will discuss a line of research that my group has been doing in the past few years, which is on API virtualization for Android. I will first describe what API virtualization is and the motivation behind the research. The motivation is technical but also has a personal twist, and I will discuss both sides. I will then describe three systems that together comprise the bulk of the research done so far---Reptor, BlueMountain, and Gesto. The first system, Reptor, is a general tool that enables API virtualization on Android. It demonstrates that enabling API virtualization on Android is feasible and practical. The other two systems, BlueMountain and Gesto, apply API virtualization to two different domains and demonstrate that API virtualization can open up new possibilities for interesting use cases. BlueMountain envisions a new storage ecosystem where innovative storage solutions can be encapsulated and distributed as Android apps. Gesto enables task automation for Android apps by allowing users to map a sequence of UI actions to a voice command or a gesture. Reptor and BlueMountain are largely "done" in the sense that they have published papers and the lead students are graduating. Gesto, on the other hand, is still ongoing.

Abstract

Efficient resource sharing among distributed jobs for training multiple deep learning models is challenging. Static resource allocation is inefficient as it often ignores the churn while the performance of dynamic resource allocation heavily depends on the scalability of a job that is unknown to the scheduler. Even worse, model developers have to handle parallel execution without enough information or expertise on available system resources. This results in poor performance scalability, which adversely affects the aggregate throughput of a cluster. We present CoDDL, an automatic resource coordinator for distributed deep learning jobs on a shared cluster. It performs two tasks. First, CoDDL automates parallelization of deep learning models when resources are determined by the scheduler. It not only frees the model developers from the burden of distributed resource management, but also achieves more efficient utilization of the resources. Second, it performs cluster-wise, dynamic resource allocation according to a pluggable policy. Our default policy, Max-Speedup, maximizes aggregate speedup by exploiting performance-resource trade-offs of deep learning jobs. Our evaluation shows that Max-Speedup improves the average job completion time by 3x over SRTF while it reduces makespan by up to 26.9x.

Abstract

Internet video streaming has experienced tremendous growth over the last few decades. However, the quality of existing video delivery critically depends on the bandwidth resource. Consequently, user quality of experience (QoE) suffers inevitably when network conditions become unfavorable. We present a new video delivery framework that utilizes client computation and recent advances in deep neural networks (DNNs) to reduce the dependency for delivering high quality video. The use of DNNs enables us to enhance the video quality independent to the available bandwidth. We design a practical system that addresses several challenges, such as client heterogeneity, interaction with bitrate adaptation, and DNN transfer, in enabling the idea. Our evaluation shows the proposed system outperforms the current state-of-the-art, enhancing the average QoE by 48.9% using the same bandwidth budget or saving 22% of bandwidth while providing the same user QoE.