Nitro: A Framework for Adaptive Code Variant Tuning

Abstract

Abstract-Autotuning systems intelligently navigate a search space of possible implementations of a computation to find the implementation(s) that best meets a specific optimization criteria, usually performance. This paper describes Nitro, a programmer-directed autotuning framework that facilitates tuning of code variants, or alternative implementations of the same computation. Nitro provides a library interface that permits programmers to express code variants along with metainformation that aids the system in selecting among the set of variants at run time. Machine learning is employed to build a model through training on this meta-information, so that when a new input is presented, Nitro can consult the model to select the appropriate variant. In experiments with five real-world irregular GPU benchmarks from sparse numerical methods, graph computations and sorting, Nitro-tuned variants achieve over 93% of the performance of variants selected through exhaustive search. Further, we describe optimizations and heuristics in Nitro that substantially reduce training time and other overheads.

Compiler-Assisted Workload Consolidation for Efficient Dynamic Parallelism on GPU

“…The 1-1 mapping configuration indicates that the kernel is configured to have as many blocks (or threads, in the case of thread-mapped child kernels) as the number of items in the buffer. The exhaustive search configuration is the best configuration we find from exhaustively searching the configuration space [16]. In Figure 6 we report the results on tree descendants for all considered consolidation granularities over two datasets.…”

On the Autotuning Potential of Time-stepping methods from Scientific Computing

“…The resulting number of code variants can be large and efficient heuristic search strategies are important, including Simulated Annealing or Nelder-Mead [31]. OpenTuner [22] provides several search strategies, but other techniques based on machine learning are also investigated [32]. Energy and performance autotuning for two irregular applications, graph community detection using the Louvain method (Grappolo) and high-performance conjugate gradient (HPCCG) have been investigated in [33] for OpenMP multithreaded programs using Open-Tuner.…”

Input-aware auto-tuning of compute-bound HPC kernels

“…Input-aware auto-tuning arose recently [10] as a way to solve this issue, and has been since then applied to a variety of problems including poly-algorithmic selection [3], OpenACC loops optimization [11], and general-purpose GPU compilers [12,14]. This surge of interest is encouraging, but has yet to win over an industry dominated by manual heuristics.…”