Threads and HID

This site is a joint effort with David Merrill and Olaf Matthes as part of our [hidio] project.

In the course of this project, we had long and involved discussions about whether we should use threads or not. We tried a number of implementations, both threaded and not. In the end, it seems that there was no difference in performance between the two. But depending on the software environment that the author is working in, threads may or may not make sense. Since we are working with Pd and Max/MSP, that was our context. Pd mostly avoids threads, while Max/MSP provides two threads.

Performance

• HID events are optimized at the kernel level to ensure they are not delayed
• two years of [hid]'s non-threaded event handling seems to be working well.
• no measured difference in latency between threading/callback vs. polling
• the OS provides an event queue, so non-blocking reads don't miss events

Implementation

• keeping event handling out of a thread can be somewhat unorthodox
• pre-emptive threads work against Pd/Max language design, can cause weirdness
• Windows can be grumpy without a thread
• Windows event reading designed to work with blocking reads in a thread
• thread would only be needed for the event polling.
• non-blocking reads are essential for threadless implementation

Examples

• CUIOSC has one implementation with a thread and one without.

$Id: threads.html,v 1.3 2007/04/14 19:15:24 hans Exp $