With the widespread use of computing devices and software in critical tasks, high software reliability and performance are paramount. To ensure that software is running reliably, it is important to monitor its execution to identify faulty behavior and then correct it. Also, since deployed software must handle a wide range of situations, it is important to monitor program's input and dependence characteristics during execution and adapt it to deliver high performance. The research project has delivered techniques to enable efficient runtime monitoring of software, tools for debugging of parallel programs to improve their reliability, and runtime techniques for improving performance of complex real-world applications.

1. Efficient Runtime Monitoring.
Continuous monitoring of running software can result in high runtime overhead, since a significant portion of the compute cycles can be taken up by monitoring activities. In this work we designed lightweight hardware support that can be programmed to perform a wide range of monitoring tasks with minimal runtime overhead. The effectiveness of programmable hardware was demonstrated by using it to perform monitoring needed for software reliability (e.g., record-and-replay) and software performance (e.g., speculative parallel execution). An important contribution of the above work is that it considers the relationship between memory models and runtime monitoring of memory accesses. We have solved the long standing problem of efficiently supporting the sequential consistency memory model which makes the task of producing reliable software much more manageable.

2. Debugging Tools for Parallel Software.
Existing debugging tools provide little guidance for the programmers towards locating the bug source. We have developed a new tool, DrDebug, for efficiently debugging multithreaded programs. By providing several new commands we make the task of examining and analyzing the state of a running program much easier. The insights gained by using these commands helps the user track down the root cause of faulty behavior and improve program understanding, allowing the programmer to modify the program to eliminate faulty behavior. DrDebug works for parallel programs and can efficiently monitor long program runs. Additional support for replaying parts of program execution is provided so that the user can efficiently explore and understand program behavior. The effectiveness of DrDebug was demonstrated by monitoring and debugging of real-world programs containing bugs.

3. Exploiting Parallelism in Real-world Applications.
Modern applications in important domains such as genomics and data mining are characterized by their need for massive computing power both due to their computational complexity and their handling of massive amounts of data. We have observed that the massive amounts of input, intermediate, and output data that these applications must handle causes programmers to develop code which continuously carries out data transfers between files and memory to make the best use of limited available memory. Frequent I/O operations that perform these data transfers introduce dependences that greatly limit our ability to exploit parallelism in hybrid loops (i.e., loops containing a mix of computation and I/O). While much research has been carried out over past several decades on exploiting parallelism, these techniques are mostly applicable to computations that are free of I/O operations. We have developed a novel technique that breaks I/O caused dependences and greatly enhances our ability to exploit parallelism for real-world applications. As an example we studied Velvet, a popular de novo genomic assembler. Velvet must deal with large input sizes and large amounts of intermediate data. For example, for processing of the wheat genome, Velvet must handle an input file that is 15 Gb in size ...

Please report errors in award information by writing to: awardsearch@nsf.gov.