Degree Type
Thesis
Date of Award
2015
Degree Name
Master of Science
Department
Computer Science
First Advisor
Hridesh Rajan
Abstract
Performance tuning is the leading justification for breaking abstraction boundaries. We target this problem for message passing concurrency (MPC) abstractions on the Java Virtual Machine (JVM). Efficient mapping of MPC abstractions to threads is critical for performance, scalability, and CPU utilization; but tedious and time consuming to perform manually. We solve this problem by putting forth a technique for automatically mapping MPC abstractions to JVM threads. In general, this mapping cannot be found in polynomial time. Our surprising observation is that characteristics of MPC abstractions and their communication patterns can be very revealing, and can help determine the mapping. Our technique addresses a number of challenges that leads to improved performance: i) balancing the computations across JVM threads, ii) reducing the communication overheads, iii) utilizing the information about cache locality, and iv) mapping MPC abstractions to threads in a way that reduces the contention between JVM threads. We have realized our technique in the Panini language that has capsules as an MPC abstraction. We also compare our mapping technique against four default mapping techniques: thread-all, round-robin-task-all, random-task-all and work-stealing. Our evaluation on wide range of benchmark programs shows that our mapping technique can improve the performance by 30%-60% over default mapping techniques.
DOI
https://doi.org/10.31274/etd-180810-3998
Copyright Owner
Ganesha Upadhyaya
Copyright Date
2015
Language
en
File Format
application/pdf
File Size
46 pages
Recommended Citation
Upadhyaya, Ganesha, "Abstraction and performance, together at last: auto-tuning message-passing concurrency on the Java virtual machine" (2015). Graduate Theses and Dissertations. 14447.
https://lib.dr.iastate.edu/etd/14447