The ability to process large amounts of data efficiently is a must in modern Big Data, Data Science and Machine Learning (ML). Easy and succinct workflow expression with highly efficient scaling is key. Existing systems such as Apache Spark and Beam were created for this type of work but have steep learning curve and tremendous operating complexity. Having a simpler tool with less overhead that is well-matched to the typical workflow definition and execution requirements of data science/ML projects will be beneficial.We propose a new framework PB&J that enables succinct definition and data-parallel scale-out execution of workflows with ease. PB&J is easy to learn as it builds on the native Unix shell, with the addition of just a few operators. PB&J supports Maximal Parallelism with its ability to do true pipelining, and fault recovery with Minimal Redo. It is well-matched to the typical data science/ML processing requirements by leveraging the existing command line executables and Python modules. We illustrate the features and strengths of PB&J with real-life Deep Learning data processing workflows. We compare it to existing frameworks such as Apache Spark, Apache Beam, Swift/T and Apache Airflow in terms of ease of authoring, efficiency, scalability and fault recovery.

Achieving good accuracy in Deep Learning (DL) models for Wake Word Engines (WWE) strongly depends on using the right type of dataset in training. Important WWE dataset dimensions are the number of wake words, the ratio of wake words to non-wake words, the composition of non-wake words and the augmentation techniques applied to the training dataset. In this paper, we make two crucial and first-of-a-kind contributions in improving the WWE performance: (a) Extensively studying and proving the effectiveness of three potent audio augmentation techniques that are previously not studied or less-studied for WWE and (b) Providing clear insight into the effect of wake-word to non-wake word ratio and the composition of the latter, specifically the impact of using non-wake words uttered by the speakers who also utter the wake words (vs) using non-wake words that are mined from elsewhere,We prove the effectiveness of our recipe of: (a) Data augmentation with Ambient Sound addition, Reverberation, Echo and Speed along with SpecAugment and (b) A non-wake word composition of Common Voice plus Freesound audio slices at a wake word to non-wake word ratio of 1:15 and 1:20 respectively. We show that the WWE model trained with our recipe for `Computer' exhibits a False Reject Rate (FRR) at 1 False Alarm (FA) per hour of 0%, 5.6%, 0.4%, and 0.7% under the four test environments trialed, whereas, the standard technique trained model exhibits an FRR of 1.1%, 18.3%, 3.7%, and 3.7%. Similar results are observed for wake word `Crystal'.

Voice interfaces are fast becoming an important human-machine interfaces, and Wake Word Engines (WWEs) are a critical part of modern voice interfaces. There are recent advancements in applying Deep Learning (DL) or Deep Neural Network architectures for WWE construction. Similar to other applications of DL however, achieving good accuracy strongly depends on training using the right type of dataset - a task that traditionally requires significant time and human effort. In this paper, we present novel techniques for curating WWE datasets that significantly minimizes the need for data collection from humans. More specifically, we investigated two techniques: (1) Using “Found Data”: we have created automated data curation pipelines for locating and processing data from public sources (e.g., YouTube), and (2) Using Synthetic Data: we explored the use of Synthesized Speech (using Text to Speech and Voice Conversion) to synthetically generate WWE datasets. Using our techniques, we are able to train WWEs that demonstrate a level of performance that is on-par with the WWEs trained with data from expensive human collection (e.g., Mechanical Turk). For example, in our experiments using wake words `Computer' and `Shannon', for a given False Alarm rate of 1 per hour, WWEs trained with our novel methods exhibit a False Reject Rate, averaged across four different test environments, of 0.9% and 1.6% respectively compared to the 1.0% and 0.8% for the baseline trained using data from human collection. Cycle time savings of more than an order of magnitude is possible by utilizing these methods.

Today's cloud storage services must offer storage reliability and fast data retrieval for large amount of data without sacrificing storage cost. We present SEARS, a cloud-based storage system which integrates erasure coding and data deduplication to support efficient and reliable data storage with fast user response time. With proper association of data to storage server clusters, SEARS provides flexible mixing of different configurations, suitable for real-time and archival applications. Our prototype implementation of SEARS over Amazon EC2 shows that it outperforms existing storage systems in storage efficiency and file retrieval time. For 3 MB files, SEARS delivers retrieval time of 2.5 s compared to 7 s with existing systems.

The cellular industry is evaluating architectures to distribute the signal processing in radio access networks. One of the options is to process the signals of all base stations on a shared pool of compute resources in a central location. In this centralized architecture, the existing base stations will be replaced with just the antennas and a few other active RF components, and the remainder of the digital processing including the physical layer will be carried out in a central location. This model has potential benefits that include a reduction in the cost of operating the network due to fewer site visits, easy upgrades, and lower site lease costs, and an improvement in the network performance with joint signal processing techniques that span multiple base stations. Further there is a potential to exploit variations in the processing load across base stations, to pool the base stations into fewer compute resources, thereby allowing the operator to either reduce energy consumption by turning the remaining processors off or reducing costs by provisioning fewer compute resources. We focus on this aspect in this paper.

Existing media providers such as YouTube and Hulu deliver videos by turning it into a progressive download. This can result in frequent video freezes under varying network dynamics. In this paper, we present CloudStream: a cloud-based video proxy that can deliver high-quality streaming videos by transcoding the original video in real time to a scalable codec which allows streaming adaptation to network dynamics. The key is a multi-level transcoding parallelization framework with two mapping options (Hallsh-based Mapping and Lateness-first Mapping) that optimize transcoding speed and reduce the transcoding jitters while preserving the encoded video quality. We evaluate the performance of CloudStream on our campus cloud testbed.

This paper proposes and studies a system, called CloudFlex, which transparently taps cloud resources to serve application requests that exceed capacity of internal infrastructure. CloudFlex operates as a feedback control system with two key interacting components: load balancer and controller. We focus on operational optimality and stability of the system, highlight the tradeoffs between cost and responsiveness, and address important design considerations such as choke point detection that are critical in avoiding pathological system operations. For evaluation, we develop a prototype of CloudFlex on our testbed comprising servers of our enterprise data center and Amazon EC2 instances.