On this page you will find a somewhat up-to-date list of the projects that I'm working on. If you'd like any additional information about any of these, feel free to e-mail me.
I decided that it would be fun to write my own VM from scratch. So that's what I did. It's a hard real-time embedded VM for safety-critical applications, and its development is currently funded by the National Science Foundation, the Indiana Economic Development Corporation, and the Air Force. Currently it's only AOT, but I'm working on fixing that. And speaking of AOT - the Fiji VM compiler is fully parallel; you can compile typical programs in around 15 seconds on a quad-core machine. The performance of the generated executables is better than Ovm and within 10% of HotSpot 1.6 server, it's got great multicore support, and it uses an Immix-like on-the-fly concurrent garbage collector with strong real-time guarantees. This collector has the world's first wait-free concurrent copying algorithm with strong coherence and throughput guarantees. It also features my usual tricks, like high performance native threading and the best locking implementation out there. Perhaps most awesome is that in addition to being able to run on POSIX-like platforms, the Fiji VM can boot from bare hardware. One of our demos is that we give people a CD with some Java programs; you can pop the CD into your x86 computer, boot from the CD, and the Java programs will run directly on the bare metal. See here for updates about this technology.
Locking is essential for object-oriented programming, especially for multithreaded and parallel programs. But sometimes locks are used unnecessarily, and are never contended. This can occur especially of locks are used declaratively. Ideally, there should be no cost to declaring a critical section as such - the VM should be able to identify dynamically if this is a real critical section (i.e. it's providing protection against simultaneous execution by multiple threads) or if it's superfluous. Biased locking is one way of enabling the VM to dynamically deactivate critical sections that aren't being contended. My implementation in Fiji VM and Jikes RVM leads to tremendous speedups even in multi-threaded programs, does not require any fancy heuristics, and has a clear worst-case bound, making it suitable in real-time programs when combined with PIP locking.
Jikes RVM has traditionally used green threads. But this is unsustainable. Real Java programs expect native thread semantics. Since April 2008, I was working on implementing native threads; this work is now in RVM trunk and has been released as part of Jikes RVM 3.1.0.
RTGC works great - if you don't have too many processors. The next step is to make it scale. We are exploring ways of scaling to 32+ processors by using a combination of reference counting, heaplets, and concurrent GC techniques.
Hardware gives us at best double-word atomic lock-free compare-and-swap (CAS). But we want to be able to do multi-word CAS. Previous techniques for emulating multi-CAS in software required stealing bits; I have a stochastic algorithm that requires no bit stealing. See here for more info.
While at Microsoft Research, working with David Tarditi, Bjarne Steensgaard, and Erez Petrank, we implemented a number of on-the-fly lock-free concurrent copying collector with strong real-time guarantees. This work was done in the Bartok research compiler. So far we have an ISMM'07, ISMM'08, and PLDI'08 paper from this work.
I just recently added a high-performance real-time garbage collector to OVM. It was quite fun. I was able to reproduce the claims of Metronome. Right now I've got an RTSS'06 paper from it as well as a LCTES'07 paper. The results were also used in my VEE'07 paper, and the collector continues to get used in our group for research. This includes a more recent RTSS paper, which used this collector infrastructure for studying the effectiveness of various scheduling techniques. See here for the presentation from RTSS. And here for the LCTES'07 presentation. Also see here for a poster I've presented about it.
We wrote a CC paper about a new way of doing accurate garbage collection when generating C++ code. See the poster here and the presentation here.
The goal is to explore the performance and configurability characteristics of asynchronous I/O inside a real time JVM. In this work, asynchronous I/O is used to implement all user-level (and some kernel-level) I/O facilities. The best description of this is in the TCES paper, for which I made a poster.
I have done a variety of work for VERITAS. My main contributions include the Harvester, which serves as the back-end DACQ; Quicklook and QLtools, which serve as a first-approximation analysis used by the observer; VBF, which is the standard data format; and all of the software for the Purdue volt-gain and afterpulsing PMT test stand.