Abstract
Shared-memory multicore architectures have become pervasive, and there is a pressing need for parallel programming models to facilitate both performance and convenience. However, most existing shared-memory programming models are tedious for programming and are prone to errors such as data race, which are difficult to debug.
To solve this problem, this thesis proposes a data race prevention scheme in the View-Oriented Parallel Programming (VOPP) paradigm. VOPP was proposed for distributed shared memory systems. It is adapted to shared-memory multicore architectures in this thesis. VOPP is a shared-memory data-centric parallel programming model, which uses views to bundle mutual exclusion with data access. In VOPP, programmers partition the shared memory into "views", which are non-overlapping sets of shared data objects. The data race prevention scheme proposed for VOPP can prevent data race through the memory protection mechanism while keeping the extra overhead low.
To improve the programmability of VOPP, this thesis proposes an automatic view access management scheme where a view is automatically acquired upon its first access, and automatically released when no longer needed, thus relieving programmers from arranging locks to protect critical sections.
To further improve performance and programmability, this thesis proposes the View-Oriented Transactional Memory (VOTM) system, which uses Restricted Admission Control (RAC) to manage the number of processes holding each view according to its contention. In VOTM, RAC can restrict the number of processes holding the view when its contention is high, and in extreme cases, RAC can fall back to the locking mode, in order to avoid abort overheads of transactions. On the other hand, RAC allows unlimited concurrent access to other low-contention views to maximize concurrency, just as in transactional memory. Therefore, VOTM has the merits of both the locking mechanism and the transactional memory (TM) and integrated them nicely through RAC.
This thesis has also provided a theoretical analysis for RAC to investigate factors that indicate performance gain by restricting admission to a view, including disproportionately large portion of time spent in aborted transactions due to high contention and excessive TM mechanism overheads. Experimental results demonstrate that in many cases, RAC correctly responds to these situations by restricting admission to a view, thus improves the performance.
Apart from the improvements of programmability in VOPP, this thesis has done extensive experiments on two multicore architectures, a 16-core machine and a 64-core machine. Experimental results demonstrate that VOPP can provide a data race free environment with low overheads on multicore architectures, and VOTM outperforms both traditional transactional memory models and lock-based models in most benchmark applications.