High-performance concurrent programming

点击翻译按钮获取中文摘要

The research objective is to provide basic and enabling technology, through programming languages, tools and the runtime environment, to exploit concurrency on shared-memory multi-threaded/core/processor computers. Computer Science has entered a new epoch where most computing devices have multiple processors, and most future programs will be written for applications running on low-cost, high-performance, commodity, parallel computers. Unfortunately, the level of understanding and necessary tools to develop this application domain is wholly inadequate for the average programmer. Without further advances, concurrent programming will remain an esoteric skill, inhibiting the full utilization and adoption of commodity parallel hardware. My research examines and extends knowledge in 5 critical areas of concurrent programming. 1. Asynchronous call implicitly introduces concurrency, and futures are used to return results from an asynchronous call. I am developing a mechanism for dynamically selecting among homogeneous and heterogeneous future types, allowing conditional selectors and "and/or" operators among selectors. 2. Asynchronous exception propagation is a useful alternative communication among threads, especially if timely propagation is ensured. However, timely propagation is impossible for blocked threads. I am developing an approach for unblocking threads to quickly begin propagation of asynchronous termination and resumption exceptions. 3. I have identified 7 fundamental memory-allocator features used to achieve good performance. Each feature is being examined in isolation to make it easier to reason about memory allocator design and provide a common base for comparison of allocators. 4. I am developing a profiling tool to aid in understanding the dynamic execution of concurrent programs running on parallel hardware to facilitate debugging and performance analysis. 5. Many concurrent applications are I/O rather that CPU intensive (web servers). I am developing approaches to utilize network and CPU parallelism to build very high-performance multi-threaded web-servers. Potential users of the research span a broad spectrum as concurrent applications are becoming crucial in many computing environments.