OpenDwarfs: Characterization of Dwarf-Based Benchmarks on Fixed and Reconfigurable Architectures

Abstract

The proliferation of heterogeneous computing platforms presents the parallel computing community with new challenges. One such challenge entails evaluating the efficacy of such parallel architectures and identifying the architectural innovations that ultimately benefit applications. To address this challenge, we need benchmarks that capture the execution patterns (i.e., dwarfs or motifs) of applications, both present and future, in order to guide future hardware design. Furthermore, we desire a common programming model for the benchmarks that facilitates code portability across a wide variety of different processors (e.g., CPU, APU, GPU, FPGA, DSP) and computing environments (e.g., embedded, mobile, desktop, server).As such, we present the latest release of OpenDwarfs, a benchmark suite that currently realizes the Berkeley dwarfs in OpenCL, a vendor-agnostic and openstandard computing language for parallel computing. Using OpenDwarfs, we characterize a diverse set of modern fixed and reconfigurable parallel platforms: multicore CPUs, discrete and integrated GPUs, Intel Xeon Phi co-processor, as well as a FPGA. We describe the computation and communication patterns exposed by a representative set of dwarfs, obtain relevant profiling data and execution information, and draw conclusions that highlight the complex interplay between dwarfs' patterns and the underlying hardware architecture of modern parallel platforms.

Characterizing Optimizations to Memory Access Patterns using Architecture-Independent Program Features

“…However, work-items accessing data along a column of the matrix do not observe memory access coalescing. us we observe that while cache hit rates are typically low on this benchmark, particularly on GPUs [12], GPUs can hide the latency of global memory accesses through memory access coalescing to some extent.…”

“…is indicates that almost all memory accesses made by the kernel are perfectly synchronized between OpenCL threads. Performance results [9,12] show that GEM performs signi cantly be er on GPUs than on CPUs, as memory unit stalls are at low levels for both CPUs and GPUs due to the highly e cient memory utilization of this benchmark. As memory operations do not present a bo leneck, this benchmark is able to take advantage of the superior oating-point compute capability of GPUs [12].…”

“…is rightly indicates that memory addresses accessed at each logical timestamp are very distant. On GPUs, this translates to poor utilization of both memory access coalescing and caching [12]. is trend in the parallel spatial locality metric suggests that a possible improvement for GPU performance would be to load blocks of the input matrix into on-chip local memory to reduce the number of global memory requests -this is typically performed in GPU optimized implementations of the Needleman-Wunsch algorithm [1].…”

“…At each point in the kernels' memory access pro le, each work-item with global ID (i, j) in an OpenCL work-group accesses the (i, j) th elements of various matrices. e sequential memory access pa ern is non-linear since di erent matrices are accessed consecutively by the kernel, prohibiting ideal caching [12]. However, memory requests made simultaneously by a work-group always fall within a rectangular block of one of the matrices.…”

Bridging the Performance-Programmability Gap for FPGAs via OpenCL: A Case Study with OpenDwarfs

“…In this paper, we use the OpenDwarfs benchmark suite [3], a suite of architecture-agnostic OpenCL kernels that capture common computation and communication patterns across a wide spectrum of scientific and engineering applications, to study the performance of the OpenCL programming model on FPGAs. In OpenDwarfs, none of the dwarfs contain optimizations that favor a specific architecture over another.…”