Nawanol Theera-Ampornpunt

Using big data for more dependability: a cellular network tale

There are many large infrastructures that instrument everything from network performance metrics to user activities. However, the collected data are generally used for long-term planning instead of improving reliability and user experience in real time. In this paper, we present our vision of how such collections of data can be used in real time to enhance the dependability of cellular network services. We first discuss mitigation mechanisms that can be used to improve reliability, but incur a high cost which prohibit them to be used except in certain conditions. We present two case studies where analyses of real cellular network traffic data show that we can identify these conditions.

Video through a crystal ball: effect of bandwidth prediction quality on adaptive streaming in mobile environments

Mobile environments are characterized by rapidly fluctuating bandwidth and intermittent connectivity. Existing video streaming algorithms can perform poorly in such network conditions because of their reactive adaptation approach. Recent efforts suggest that bitrate adaptation using proactive accurate bandwidth prediction can help improve the quality of experience (QoE) of video streaming. However, highly accurate long-term predictions may be needed in mobile environments and those can be difficult to obtain. In this work, we examine the impact of bandwidth prediction quality on the QoE. We first characterize bandwidth profiles where bandwidth prediction-based adaptation can be useful. We then study the impact of prediction horizon and errors on the performance of Adaptive Bitrate (ABR) streaming. We observe that performance improves as the prediction horizon increases at first and then benefits start to diminish. We demonstrate that with proper error mitigation heuristic, even erroneous predictions can be useful in some scenarios. Finally, we study the role of video system parameters, namely buffer size and bitrate granularity on bandwidth prediction-based adaptation.

Automatic Problem Localization via Multi-dimensional Metric Profiling

Debugging todays large-scale distributed applications is complex. Traditional debugging techniques such as breakpoint-based debugging and performance profiling require a substantial amount of domain knowledge and do not automate the process of locating bugs and performance anomalies. We present Orion, a framework to automate the problem-localization process in distributed applications. From a large set of metrics, Orion intelligently chooses important metrics and models the applications runtime behavior through pair wise correlations of those metrics in the system, within multiple non-overlapping time windows. When correlations deviate from those of a learned correct model due to a bug, our analysis pinpoints the metrics and code regions (class and method within it) that are most likely associated with the failure. We demonstrate our framework with several real-world failure cases in distributed applications such as: HBase, Hadoop DFS, a campus-wide Java application, and a regression testing framework from IBM. Our results show that Orion is able to pinpoint the metrics and code regions that developers need to concentrate on to fix the failures.

A study of soft error consequences in hard disk drives

Hard disk drives have multiple layers of fault tolerance mechanisms that protect against data loss. However, a few failures occasionally breach the entire set of mechanisms. To prevent such scenarios, we rely on failure prediction mechanisms to raise alarms with sufficient warning to allow the at-risk data to be copied to a safe location. A common failure prediction technique monitors the occurrence of soft errors and triggers an alarm when the soft error rate exceeds a specified threshold. This study uses data collected from a population of over 50,000 customer deployed disk drives to examine the relationship between soft errors and failures, in particular failures manifested as hard errors. The data analysis shows that soft errors alone cannot be used as a reliable predictor of hard errors. However, in those cases where soft errors do accurately predict hard errors, sufficient warning time exists for preventive actions.

Fast training on large genomics data using distributed Support Vector Machines

The field of genomics has seen a glorious explosion of high-quality data, with tremendous strides having been made in genomic sequencing instruments and computational genomics applications meant to make sense of the data. A common use case for genomics data is to answer the question if a specific genetic signature is correlated with some disease manifestations. Support Vector Machine (SVM) is a widely used classifier in computational literature. Previous studies have shown success in using these SVMs for the above use case of genomics data. However, SVMs suffer from a widely-recognized scalability problem in both memory use and computational time. It is as yet an unanswered question if training such classifiers can scale to the massive sizes that characterize many of the genomics data sets. We answer that question here for a specific dataset, in order to decipher whether some regulatory module of a particular combinatorial epigenetic “pattern” will regulate the expression of a gene. However, the specifics of the dataset is likely of less relevance to the claims of our work. We take a proposed theoretical technique for efficient training of SVM, namely Cascade SVM, create our classifier called EP-SVM, and empirically evaluate how it scales to the large genomics dataset. We implement Cascade SVM on the Apache Spark platform and open source this implementation1. Through our evaluation, we bring out the computational cost on each application process, the way of distributing the overall workload among multiple processes, which can potentially execute on different cores or different machines, and the cost of data transfer to different cores or different machines. We believe we are the first to shed light on the computational and network costs of training an SVM on a multi-dimensional genomics dataset. We also evaluate the accuracy of the classifier result as a function of the parameters of the SVM model.

Interpretable deep neural networks for enhancer prediction

Enhancers are short DNA sequences that modulate gene expression patterns. Recent studies have shown that enhancer elements could be enriched for certain histone modification combinatorial codes, leading to interest in developing computational models to predict enhancer locations. Here we present EP-DNN, a protocol for predicting enhancers based on chromatin features, in two different cell types, a human embryonic (H1) and a human lung fibroblast (IMR90) cell line. Specifically, we use a deep neural network (DNN)-based architecture to extract enhancer signatures. We train EP-DNN using distal p300 binding sites, as enhancers, and TSS and random non-DNase-I hypersensitivity sites, as non-enhancers. We find that EP-DNN has superior accuracy relative to other state-of-the-art algorithms, such as DEEP-EN and RFECS, and also scales well to large number of predictions. Then, we surmount the problem that DNN results are not interpretable and develop a method to interpret which histone modifications are important, and within that, which spatial features proximal or distal to the enhancer site, are important. We uncover that the important histone modifications vary between cell types. Further, whether the important features are clustered around the enhancer peak or more spread out also differs among the different histone modifications. Thus, we bring forth a new paradigm for automatically determining the important features and the important histone modifications, rather than the current computational standard of using the same fixed number of features from all the histone modifications for all cell types. Our results have implications for computational scientists who can now do feature selection for their classification task and for biologists who can now experimentally collect data only for the relevant histone modifications.

Sense-aid: a framework for enabling network as a service for participatory sensing

The rapid adoption of smartphones with different types of advanced sensors has led to an increasing trend in the usage of mobile crowdsensing applications, e.g., to create hyper-local weather maps. However, the high energy consumption of crowdsensing, chiefly due to expensive network communication, has been found to be detrimental to the wide-spread adoption. We propose a framework, called Sense-Aid, that can provide energy-efficient mobile crowdsensing service, coexisting with the cellular network. There are two key innovations in Sense-Aid beyond prior work (Piggyback Crowdsensing-Sensys13)---the middleware running on the cellular network edge to orchestrate multiple devices present in geographical proximity to suppress redundant data collection and communication. It understands the state of each device (radio state, battery state, etc.) to decide which ones should be selected for crowdsensing activities at any point in time. It also provides a simple programming abstraction to help with the development of crowdsensing applications. We show the benefit of Sense-Aid by conducting a user study consisting of 60 students in our campus, compared to a baseline periodic data collection method and Piggyback Crowdsensing. We find that energy saving is 93.3% for Sense-Aid compared with Piggyback Crowdsensing in a representative case which requires 2 devices to provide barometric values within an area of a circle whose radius is 1 kilometer and requires periodic data collection every 5 minutes for a 90-minute test. The selection algorithm of Sense-Aid also ensures reasonable fairness in the use of the different devices.

TANGO: Toward a More Reliable Mobile Streaming through Cooperation between Cellular Network and Mobile Devices

Multimedia streaming is a major mobile application, accounting for more than half of total mobile traffic. Streaming applications usually have a static buffering strategy. For example, buffer size is limited to x minutes of the stream, where x is optimized to provide the best trade-off between minimizing stalls and limiting waste of users bandwidth and energy resulting from user abandonment. We show that such strategies based on information available on the mobile device alone do not work well when network conditions change dynamically, e.g., connectivity degrades due to congestion. We propose an alternative strategy using the framework called TANGO, based on a novel idea of cooperation between cellular network and mobile devices. By monitoring real-time network conditions and continuously predicting user location, our system is able to predict connectivity degradation in the near term. In such events, a notification is sent to the mobile device so that the streaming application can initiate a mitigation action, such as to pre-cache more content. In simulations based on real user traces, we found that TANGO reduces pause time by 13–72%, significantly outperforming DASH, which is the current state of the art.