Accelerating Data-intensive Bioinformatic Workloads

Restricted (Penn State Only)
- Author:
- Li, Zheyu
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- November 16, 2023
- Committee Members:
- Chitaranjan Das, Program Head/Chair
Vijaykrishnan Narayanan, Chair & Dissertation Advisor
Abhronil Sengupta, Major Field Member
John Sampson, Major Field Member
Santhosh Girirajan, Outside Unit & Field Member - Keywords:
- Bioinformatic
Genome sequencing
FPGA
In-memory computing
Whole Silde Imaging
Max Cut
Nanopore Sequencing
Sorting - Abstract:
- Bioinformatic workloads such as Whole Slide Imaging, and genome sequencing emerged rapidly in recent years. These workloads often utilize data-intensive operations such as combinatorial optimizations, sorting, and deep neural network inferences. However, enormous computation resources and data movements are often required to process the collected biological data, which limits the practical usage of these methods. We aim to enable efficient and high-performance processing of bioinformatic workloads utilizing both algorithm designs which optimize the workflow pattern, and hardware designs which leverage advanced compute device and memory technologies. Genome sequencing data can contain up to hundreds of million reads. Identification of certain genetic modifications thus requires sorting a large number of reads which leads to significant process time. In our first work, we proposed IMC-Sort, a Process-In-Memory(PIM) architecture that leverages Hybrid-Memory-Cube(HMC) technology. The main idea is to augment the logic layer of the HMC device thus the memory device itself can support specific data operations. By offloading the sorting operations to the memory devices, the internal bandwidth can be efficiently exploited and expensive data movement can be reduced. Whole Slide Image (WSI) analysis is increasingly being adopted as an important tool in modern pathology. Most of the existing analysis approaches require the complete decompression of the whole image exhaustively, which leads to inefficient processing. In our second work, we present compression domain processing-based computation efficient analysis workflows for WSIs classification which leverages the pyramidal magnification structure of WSI files and compression domain features that are available from the raw code stream. Different decompression depths are assigned to the high-magnification level patches at different locations at fine-grained, thus avoiding exhaustive full decompression. Many genome analyses can be mapped to combinatorial problems, such as haplotype assembly which is based on computing max-cuts in certain graphs derived from the sequenced fragments. In our third work, we demonstrate a one of its kind Field-Programmable-Gate-Array(FPGA)-based compute engine that uses new computational models inspired by the synchronization dynamics of oscillators to solve the general form of the computationally intractable Max-K-Cut combinatorial optimization problem which can be applied to genome analyses. Targeting the 3rd generation of DNA sequencing technologies such as Nanopore, Adaptive Banded Event Alignment (ABEA) is a crucial algorithmic component of polishing, aligning the raw signal to a reference sequence via a dynamic programming strategy. In our fourth work, we proposed a FPGA-based hardware accelerator that significantly improved the throughput and energy efficiency. In summary, targeting data-intensive bioinformatic workloads, this thesis presents Compressed-domain-processing methodologies to reduce compute/memory overheads and hardware designs to exploit fine-grained parallelism and improve energy efficiencies.