Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical and Computer Engineering

First Advisor

Saurabh Bagchi

Committee Chair

Saurabh Bagchi

Committee Member 1

Milind Kulkarni

Committee Member 2

Samuel P. Midkiff

Committee Member 3

Jennifer Neville


Detection, diagnosis and mitigation of performance problems in today's large-scale distributed and parallel systems is a difficult task. These large distributed and parallel systems are composed of various complex software and hardware components. When the system experiences some performance or correctness problem, developers struggle to understand the root cause of the problem and fix in a timely manner. In my thesis, I address these three components of the performance problems in computer systems. First, we focus on diagnosing performance problems in large-scale parallel applications running on supercomputers. We developed techniques to localize the performance problem for root-cause analysis. Parallel applications, most of which are complex scientific simulations running in supercomputers, can create up to millions of parallel tasks that run on different machines and communicate using the message passing paradigm. We developed a highly scalable and accurate automated debugging tool called PRODOMETER, which uses sophisticated algorithms to first, create a logical progress dependency graph of the tasks to highlight how the problem spread through the system manifesting as a system-wide performance issue. Second, uses this logical progress dependence graph to identify the task where the problem originated. Finally, PRODOMETER pinpoints the code region corresponding to the origin of the bug. Second, we developed a tool-chain that can detect performance anomaly using machine-learning techniques and can achieve very low false positive rate. Our input-aware performance anomaly detection system consists of a scalable data collection framework to collect performance related metrics from different granularity of code regions, an offline model creation and prediction-error characterization technique, and a threshold based anomaly-detection-engine for production runs. Our system requires few training runs and can handle unknown inputs and parameter combinations by dynamically calibrating the anomaly detection threshold according to the characteristics of the input data and the characteristics of the prediction-error of the models. Third, we developed performance problem mitigation scheme for erasure-coded distributed storage systems. Repair operations of the failed blocks in erasure-coded distributed storage system take really long time in networked constrained data-centers. The reason being, during the repair operation for erasure-coded distributed storage, a lot of data from multiple nodes are gathered into a single node and then a mathematical operation is performed to reconstruct the missing part. This process severely congests the links toward the destination where newly recreated data is to be hosted. We proposed a novel distributed repair technique, called Partial-Parallel-Repair (PPR) that performs this reconstruction in parallel on multiple nodes and eliminates network bottlenecks, and as a result, greatly speeds up the repair process. Fourth, we study how for a class of applications, performance can be improved (or performance problems can be mitigated) by selectively approximating some of the computations. For many applications, the main computation happens inside a loop that can be logically divided into a few temporal segments, we call phases. We found that while approximating the initial phases might severely degrade the quality of the results, approximating the computation for the later phases have very small impact on the final quality of the result. Based on this observation, we developed an optimization framework that for a given budget of quality-loss, would find the best approximation settings for each phase in the execution.