Fun and Interesting Thoughts (FIT) is a special event held at the 2010 Conference on Programming Language Design and Implementation (PLDI). FIT is a venue for thoughts that present fresh insights, unveil surprising ideas, identify hidden trends, or that are simple neat and fun for a PLDI audience.
Monday, May 24, 2010
Outfoxing the Mammoths
by Marek Olszewski and Saman Amarasinghe, Massachusetts Institute of Technology
Some of last year’s FIT talks motivated and provided uses for highly sought-after features for tomorrow’s programming languages, such as deterministic parallel programs and ultra fast always-on precise race detectors. We argue that if we want to solve these daunting problems we need additional ammunition from the architecture and operating system communities. However, operating systems and computer architectures are mammoth systems that are changing too slowly for us to rely on. Thus, rather than waiting and hoping to receive the ammunition, we need to outfox them so that we can obtain the tools we need today. In this paper, we suggest using hypervisors to augment the system stack with new features that can help researchers develop new systems. By using this approach, which operates transparently with existing systems, we will be better able to spur the mammoths into action. As an example of such a hypervisor, we have developed a hypervisor that provides threads, within a single address space, the ability to set per-thread page protections, which can be used to cheaply detect sharing patterns between threads at a coarse granularity. We hope this feature will allow new work in the areas of race detection, deterministic multithreading and software transactional memory.
Fun and Interesting Thoughts (FIT) is a special event held at the 2010 Conference on Programming Language Design and Implementation (PLDI). FIT is a venue for thoughts that present fresh insights, unveil surprising ideas, identify hidden trends, or that are simple neat and fun for a PLDI audience.
Some of last year’s FIT talks motivated and provided uses for highly sought-after features for tomorrow’s programming languages, such as deterministic parallel programs and ultra fast always-on precise race detectors. We argue that if we want to solve these daunting problems we need additional ammunition from the architecture and operating system communities. However, operating systems and computer architectures are mammoth systems that are changing too slowly for us to rely on. Thus, rather than waiting and hoping to receive the ammunition, we need to outfox them so that we can obtain the tools we need today. In this paper, we suggest using hypervisors to augment the system stack with new features that can help researchers develop new systems. By using this approach, which operates transparently with existing systems, we will be better able to spur the mammoths into action. As an example of such a hypervisor, we have developed a hypervisor that provides threads, within a single address space, the ability to set per-thread page protections, which can be used to cheaply detect sharing patterns between threads at a coarse granularity. We hope this feature will allow new work in the areas of race detection, deterministic multithreading and software transactional memory.
ReplyDeleteCongratulations for winning the best talk!!
ReplyDelete