Manoj P. Saha

Ph.D. Candidate


Lecture Information:
  • March 8, 2023
  • 10:00 AM
  • Zoom

Speaker Bio

Manoj P. Saha is a Ph.D. Candidate working with Prof. Janki Bhimani at the Knight Foundation School of Computing & Information Sciences (KFSCIS) at Florida International University (FIU) in the Data Management Research Lab (DaMRL). His research focuses on flash-based storage systems, persistent memory, and deep learning infrastructure. Manoj has published one peer-reviewed article as first author and three peer-reviewed articles as co-author in conferences such as Design Automation Conference (DAC), ACM (Association for Computing Machinery) International Conference on Systems and Storage (SYSTOR), and the International Performance Computing and Communications Conference (IPCCC). He also holds a U.S. patent on systems and methods for optimizing data management within key value storage. Prior to joining DaMRL, Manoj earned a MS in Computer Science from The University of Texas at El Paso, an MBA in Marketing from the Institute of Business Administration at the University of Dhaka, and a BS in Electronics and Telecommunication from North South University. He has worked in telecommunications, digital media, and marketing industries for seven years.

Description

The explosion of unstructured data has given rise to Key-Value (KV) and object stores. However, conventional operating system (OS) I/O stack incur additional computational and memory overheads impacting the performance of the software KV/object databases. I/O requests go through multiple layers of the data packing and indexing, and each layer adds syntax translation complexity and performance cost. Upon scaling the number of storage devices to meet increasing capacity demands, the indexing and data translation bottleneck of the I/O stack intensifies. Therefore, we need new methods and algorithms to offload indexing tasks from host to storage devices. This will reduce the centralized bottleneck of data indexing on host to manage data from multiple devices as well as allow the KV/object databases to scale better for increasing storage requirements. To better understand the benefits and bottlenecks of the proposed design, we first conduct a systematic study of state of art Key-Value Solid State Drive (KV-SSD) design and their internal components. However, there is a void for a non-expensive and extensible research platform that enables in-depth exploration of the index management components within the KV device. Hence, second, we design Modular Key-value Emulator (MoKE), a software emulator for fostering future full-stack software/hardware KV and object storage device research. Finally, we explore how persistent memory can be leveraged to accelerate deep learning training. In this dissertation proposal, we leverage the key benefits of emerging technologies to improve evolving application performance. Next, we want to explore if the semantic information embedded in keys can be utilized to improve SSD performance. In addition, we plan to perform in-depth exploration to understand the needs of the evolving Deep Learning workloads and design efficient systems to meet its requirements. applications.