Search for: All records

This work presents the first-ever detailed and large-scale measurement analysis of storage consumption behavior of applications (apps) on smart mobile devices. We start by carrying out a five-year longitudinal static analysis of millions of Android apps to study the increase in their sizes over time and identify various sources of app storage consumption. Our study reveals that mobile apps have evolved as large monolithic packages that are packed with features to monetize/engage users and optimized for performance at the cost of redundant storage consumption. We also carry out a mobile storage usage study with 140 Android participants. We built and deployed a lightweight context-aware storage tracing tool, called cosmos, on each participant's device. Leveraging the traces from our user study, we show that only a small fraction of apps/features are actively used and usage is correlated to user context. Our findings suggest a high degree of app feature bloat and unused functionality, which leads to inefficient use of storage. Furthermore, we found that apps are not constrained by storage quota limits, and developers freely abuse persistent storage by frequently caching data, creating debug logs, user analytics, and downloading advertisements as needed. Finally, drawing upon our findings, we discuss the need for efficient mobile storage management, and propose an elastic storage design to reclaim storage space when unused. We further identify research challenges and quantify expected storage savings from such a design. We believe our findings will be valuable to the storage research community as well as mobile app developers. 

We describe the design, deployment and operation of a computer system built to efficiently run deep learning frameworks. The system consists of 16 IBM POWER9 servers with 4 NVIDIA V100 GPUs each, interconnected with Mellanox EDR InfiniBand fabric, and a DDN all-flash storage array. The system is tailored towards efficient execution of the IBM Watson Machine Learning enterprise software stack that combines popular open-source deep learning frameworks. We build a custom management software stack to enable an efficient use of the system by a diverse community of users and provide guides and recipes for running deep learning workloads at scale utilizing all available GPUs. We demonstrate scaling of a PyTorch and TensorFlow based deep neural networks to produce state-of-the-art performance results. 

Recent advances in cyber-infrastructure have enabled digital data sharing and ubiquitous network connectivity between scientific instruments and cloud-based storage infrastructure for uploading, storing, curating, and correlating of large amounts of materials and semiconductor fabrication data and metadata. However, there is still a significant number of scientific instruments running on old operating systems that are taken offline and cannot connect to the cloud infrastructure, due to security and network performance concerns. In this paper, we propose BRACELET - an edge-cloud infrastructure that augments the existing cloud-based infrastructure with edge devices and helps to tackle the unique performance & security challenges that scientific instruments face when they are connected to the cloud through public network. With BRACELET, we put a networked edge device, called cloudlet, in between the scientific instruments and the cloud as the middle tier of a three-tier hierarchy. The cloudlet will shape and protect the data traffic from scientific instruments to the cloud, and will play a foundational role in keeping the instruments connected throughout its lifetime, and continuously providing the otherwise missing performance and security features for the instrument as its operating system ages.