This talk will be divided into two parts.
The first part will address a traditional cactus stack problem in parallel programming systems such as MIT Cilk. A cactus stack is distributed and shared among multiple threads. To avoid unbounded space problem, MIT Cilk allocated shadow frames from heap and adopted source translation to implement stack sharing. However, this approach prevented MIT legacy code from calling back Cilk code. To solve this problem, the Thread Private Storage (TPS) is proposed and implemented in Linux threads so that Cilk code can interact with legacy code freely while space and time bounds are guaranteed. This work was carried out while Zhiyi was staying at MIT Cilk group as a visiting scientist this year.
The second part of the talk will address a traditional data race issue. Data races are notoriously known as hard to debug. Many schemes are proposed to detect data races, however, they are neither inefficient (dozens times slower) or insufficient (can't detect all races). Our parallel programming system, Maotai 3.0, which is based on VOPP, has adopted a data race prevention scheme. It can fully prevent data races and generate runtime faults when they happen. In addition, Maotai 3.0 can automatically detect view access so that programmers do not need to explicitly acquire view with acquire_view(). This interface has greatly simplified parallel programming and sequential programs can be easily adapted to parallel execution with minimal modifications. However, there are implications for programmers in terms of data sharing. This talk will spell out the new protocols for the programmers. The work was carried out by Kai-Cheung and remotely supervised by Zhiyi during his sabbatical leave.
Last modified: Tuesday, 22-Sep-2009 08:08:24 NZST
This page is maintained by the seminar list administrator.