National and Kapodistrian University of Athens
A GPU performance estimation model based on micro-benchmarks and black-box kernel profiling
Abstract
dc:descriptionOver the last decade GPUs have been established in the High Performance Computing sector as compute accelerators. The primary characteristics that justify this modern trend are the exceptionally high compute throughput and the remarkable power efficiency of GPUs. However, GPU performance is highly sensitive to many factors, e.g. the type of memory access patterns, branch divergence, the degree of parallelism and potential latencies. Consequently, the execution time of a kernel on a GPU is a difficult to predict measure. Unless the kernel is latency bound, a rough estimate of the execution time on a particular GPU could be provided by applying the roofline model, which is used to map the program’s operation intensity to the peak expected performance on a particular processor.Though this approach is straightforward, it cannot not provide accurate prediction results. In this thesis, after validating the roofline principle on GPUs by employing a micro-benchmark, an analytical throughput oriented performance model is proposed. In particular, this improves on the roofline model following a quantitative approach and a completely automated GPU performance prediction technique is presented. In this respect, the proposed model utilizes micro-benchmarking and profiling in a “black-box” fashion as no inspection of source/binary code is required. The proposed model combines GPU and kernel parameters in order to characterize the performance limiting factor and to predict the execution time on target hardware, by taking into account the efficiency of beneficial computational instructions. In addition, the “quadrant-split” visual representation is proposed, which captures the characteristics of multiple processors in relation to a particular kernel.The experimental evaluation combines test executions on stencil computations (red/black SOR, LMSOR), matrix multiplication (SGEMM) and a total of 28 kernels of the Rodinia benchmark suite, all applied on six CUDA GPUs. The observed absolute error in predictions was 27.66% in the average case. Special cases of mispredicted results were investigated and justified. Moreover, the aforementioned micro-benchmark was used as a subject for performance prediction and the exhibited results were very accurate. Furthermore, the performance model was also examined in a cross vendor configuration by applying the prediction method on the same kernel source codes through the HIP programming environment supported on the AMD ROCm platform. Prediction errors were comparable to CUDA experiments despite the significant architectural differences evident between different vendor GPUs.
Degree
thesis:*- Grantor dc:publisher
- National and Kapodistrian University of Athens
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Konstantinidis, Elias
- Κωνσταντινίδης, Ηλίας
Subjects
dc:subject × 10Rights
dc:rights- Statement dc:rights
-
- BY_SA
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/41390