The evolution of big data processing and analysis has led to data-parallel frameworks such as Hadoop, MapReduce, Spark, and Hive, which are capable of analyzing large streams of data such as server logs, web transactions, and user reviews. Videos are one of the biggest sources of data and dominate the Internet traffic. Video processing on a large scale is critical and challenging as videos possess spatial and temporal features, which are not taken into account by the existing data-parallel frameworks. There are a broad range of users who want to apply sophisticated video processing pipelines such as transcoding, feature extraction, classification, scene cut detection and digital compositing to video content

Parallel video processing poses several significant research challenges to the existing data processing frameworks. Current systems are capable of processing videos but with higher resource startup times, a small degree of parallelism, low average resource utilization, coarse-grained billing, and higher latency. This research proposes a low latency software run-time for processing a single video efficiently by orchestrating cloud-based microservices. The system leverages lightweight microservices provided by Amazon Web Services Lambda framework.

This thesis explores, from both a systems and security perspective, a relatively new serverless cloud computing model that runs on stateless container invocations. This thesis demonstrates that the responsiveness, concurrency, and cost of the serverless cloud computing model, on the one hand, enables making design decisions that were not possible before, yet on the other hand, introduces security vulnerabilities that potentially increase the attack surface due to a high probability of achieving co-residence.

This thesis extends a serverless video processing framework for the cloud, Sprocket, that takes advantage of the container computing model in its design in order to provide runtime performance, resource efficiency, and a highly configurable developing environment. Sprocket demonstrates all three qualities when running both a facial recognition and streaming video pipeline. Sprocket can be configured to process a 30-minute video 1000-way parallel in under a minute for less than $1. This thesis also presents an increased attack surface in the container service that Sprocket runs on, AWS Lambda, and shows how to detect co-resident containers on physical machines across victim and adversary accounts. Co-residency is achieved in three unique Amazon data centers, with as many as 30 co-resident containers during one 30-second “attack.” This thesis demonstrates that the scalability of the serverless cloud computing model, in particular, inherently increases the probability of co-residency and that existing mitigations for co-residency must be re-evaluated.

We present MATE, a new model for developing communication-tolerant scientific applications. MATE employs a combination of mechanisms to reduce or hide the cost of network and intra-node communication. While previous approaches have been proposed to either source of communication overhead separately, the contribution of MATE is demonstrating the symbiotic effect of reducing both forms of data movement taken together in a single unified model. We explain the rationale behind our model and show its effectiveness in three scientific computing motifs on up to 64k cores of the NERSC Cori supercomputer. Lastly, we show how MATE can improve the workload balance of an irregular multigrid solver.