CPC 2006 CPC 2006

Call for Participation
Current List of
		      Participants
Program (List of Talks)
Registration Form
Accomodation
Travel Information
Links

Funded by the Vicerreitoría de Investigación of the University of A Coruña and the Ministry of Education and Science of Spain

Universidade de A Coruña Ministerio de
		      Educación y Ciencia



The Mitosis Compiler: Speculative Parallelization using Pre-Computation

Carlos García Quiñones1, Carlos Madriles1, Pedro Marcuello1, Antonio González1,2 and Dean Tullsen3
1 Intel Barcelona Research Center, Barcelona, Spain
2 Department of Computer Architecture, Universitat Politcénica de Catalunya, Barcelona, Spain
3 Department of Computer Science and Engineering, UC San Diego, USA


Speculative parallelization can provide significant sources of additional thread-level parallelism, especially for irregular applications that are hard to parallelize by conventional approaches. In this paper, we present the Mitosis compiler, which partitions applications into speculative threads, with special emphasis on applications for which conventional parallelizing approaches fail.

The management of inter-thread data dependences is crucial for the performance of the system. The Mitosis framework uses a pure software approach to predict/compute the threadŐs input values. This software approach is based on the use of pre-computation slices (p-slices), which are built by the Mitosis compiler and added at the beginning of the speculative thread. P-slices must compute thread input values accurately but they do not need to guarantee correctness, since the underlying architecture can detect and recover from misspeculations. This allows the compiler to use aggressive/unsafe optimizations to significantly reduce their overhead. The most important optimizations included in the Mitosis compiler and presented in this paper are branch pruning, memory and register dependence speculation, and early thread squashing.

Performance evaluation of the Mitosis compiler/architecture has been evaluated for a set of the Olden benchmarks. These programs have been shown to be very hard to autoparallelize by a state-of-the-art compiler. The main reason is that these codes are full of pointers and dynamically-linked data structued which make the task of the autoparalelizer really tough. Results obtained by the Mitosis compiler shows an average speedup of 2.2 when four thread units are assumed.


Back to the Workshop Program 

Please contact our webadmin with any comments or changes.