A Performance Model for Memory Bandwidth Constrained Applications on Graphics Engines

Abstract

Abstract-Graphics engines are excellent execution platforms for high-throughput computations that exploit a large degree of available parallelism. The achieved performance is, however, highly dependent on the access patterns that the application imposes on the memory subsystem. Here, we propose an analytic model that helps improve the understanding of the performance of memory-limited kernels that employ random memory access schemes, especially as impacted by cache and various configuration parameters that can be used to tune kernel execution, such as the number of blocks and the number of threads per block. The analytic model is first explored through the use of a synthetic micro-benchmark, which is then followed by an empirical validation using a pair of production applications used in computational biology.

Performance modeling for highly-threaded many-core GPUs

“…The model presented in [19], which we will extend in the subsection that follows, characterizes algorithm performance in terms of the following factors: algorithmic complexity, f app , caching, f cache , and scheduling, f sched . The algorithmic complexity factor is expressed via a function…”

“…If B r > B a × P/Q, multiple passes are needed to consume all the requested blocks of work. From [19], the number of active blocks B a is described in equation (4) in terms of the shared memory required by the application S B , the quantity of shared memory on each multiprocessor Z, the number of threads requested per thread block T r , the processor registers required by the application R T ×T r , the quantity of registers available per multiprocessor R, the maximally allowed thread blocks B max , and the maximally allowed threads T maxM P .…”

“…In the present work, we utilize both asymptotic analysis and calibrated performance prediction on many-core GPUs, in effect drawing on the concepts of [10] and [19]. We develop an integrated analytical framework combining both, analyzing algorithm efficiency and predicting the achievable execution time based on a quantification of parallelism, latency-hiding, and occupancy.…”

“…These four metrics help to identify performance bottlenecks and suggest what types of optimizations should be done. Ma et al [2], [19] design an analytic model for memory-limited kernels especially as impacted by cache and various configuration parameters that can be used to tune kernel execution. This model also reflects warp scheduling effect at the thread block level rather than instruction level, thus it is simpler than the model in [18].…”

Analysis of classic algorithms on GPUs

“…A number of highperformance GPU algorithms have been developed, such as sorting [1], hashing [2], dynamic programming [3], graph algorithms [4], and other classic algorithms [5]. Many performance studies have also been conducted [6], [7] to understand the performance of GPU applications.…”