Optimistic parallelism benefits from data partitioning

Abstract

Recent studies of irregular applications such as finite-element mesh generators and data-clustering codes have shown that these applications have a generalized data parallelism arising from the use of iterative algorithms that perform computations on elements of worklists. In some irregular applications, the computations on different elements are independent. In other applications, there may be complex patterns of dependences between these computations.The Galois system was designed to exploit this kind of irregular data parallelism on multicore processors. Its main features are (i) two kinds of set iterators for expressing worklist-based data parallelism, and (ii) a runtime system that performs optimistic parallelization of these iterators, detecting conflicts and rolling back computations as needed. Detection of conflicts and rolling back iterations requires information from class implementors.In this paper, we introduce mechanisms to improve the execution efficiency of Galois programs: data partitioning, data-centric work assignment, lock coarsening, and over-decomposition. These mechanisms can be used to exploit locality of reference, reduce mis-speculation, and lower synchronization overhead. We also argue that the design of the Galois system permits these mechanisms to be used with relatively little modification to the user code. Finally, we present experimental results that demonstrate the utility of these mechanisms.

Enhanced speculative parallelization via incremental recovery

“…The FailedAt attribute of t indicates which live-in variables caused the speculation to fail. Since, when each live-in variable is first read, the instruction address of the read and the space ID before the read are recorded (Figure 10, line 2), the main thread can retrieve these two values of the first accessed live-in variable by calling another two auxiliary functions GetRecoveryPC and getRecoverySpaceID (lines [16][17]. These two values determine the starting point of the reexecution.…”

“…Instead of state separation, Kulkarni et al proposed a rollback based speculative parallelization technique [15][16][17][18]21]. They introduce two special constructs that users can employ to identify speculative parallelism.…”

Effective parallelization of loops in the presence of I/O operations

“…Partitioned global address space (PGAS) [6,7] is a set of parallel programming models which aim to combine the performance advantage of MPI with the programmability of a shared-memory model. Galois [14,15] introduces a programming model to exploit the data parallelism in irregular applications. SpiceC [10] is a recently proposed parallel programming model for both multicores and manycores.…”

Structure-driven optimizations for amorphous data-parallel programs

“…Each partition has a lock that must be acquired by a thread that wants to access a node or edge in that partition. The over-decomposition permits a thread to do useful work even if one of its partitions is temporarily locked by a different thread [15]. It is difficult to partition the underlying graph in the Boruvka MST algorithm since it is ultimately coalesced into a single node, so we do not use partitioning for this algorithm; instead we associate locks with nodes and edges of the graph.…”