Towards Collaborative Performance Tuning of Algorithmic Skeletons

Chris Cummins, Pavlos Petoumenos, Michel Steuwer, Hugh Leather

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract / Description of output

The physical limitations of microprocessor design have forced the industry towards increasingly heterogeneous designs to extract performance. This trend has not been matched with adequate software tools, leading to a growing disparity between the availability of parallelism and the ability for application developers to exploit it.

Algorithmic skeletons simplify parallel programming by providing high-level, reusable patterns of computation. Achieving performant skeleton implementations is a difficult task; skeleton authors must attempt to anticipate and tune for a wide range of architectures and use cases. This results in implementations that target the general case and cannot provide the performance advantages that are gained from tuning low level optimization parameters. Autotuning combined with machine learning offers promising performance benefits in these situations, but the high cost of training and
lack of available tools limits the practicality of autotuning for real world programming. We believe that performing autotuning at the level of the skeleton library can overcome these issues.

In this work, we present OmniTune — an extensible and distributed framework for dynamic autotuning of optimization parameters at runtime. OmniTune uses a client-server model with a flexible API to support machine learning enabled autotuning. Training data is shared across a network of cooperating systems, using a collective approach to performance tuning.

We demonstrate the practicality of OmniTune in a case study using the algorithmic skeleton library SkelCL. By automatically tuning the workgroup size of OpenCL Stencil skeleton kernels, we show that that static tuning across a range of GPUs and programs can achieve only 26% of the optimal performance, while OmniTune achieves 92% of this maximum, equating to an average 5.65× speedup. OmniTune achieves this without introducing a significant runtime overhead, and enables portable, cross-device and cross-program tuning.
Original languageEnglish
Title of host publicationWorkshop on High-Level Programming for Heterogeneous and Hierarchical Parallel Systems (HLPGPU) 2016 @ HiPEAC, Prague, Czech Republic, January 19, 2016
Number of pages10
Publication statusPublished - 19 Jan 2016
EventHigh-Level Programming for Heterogeneous and Hierarchical Parallel Systems 2016 - Prague, Czech Republic
Duration: 19 Jan 201619 Jan 2016


ConferenceHigh-Level Programming for Heterogeneous and Hierarchical Parallel Systems 2016
Abbreviated titleHLPGPU 2016
Country/TerritoryCzech Republic
Internet address


Dive into the research topics of 'Towards Collaborative Performance Tuning of Algorithmic Skeletons'. Together they form a unique fingerprint.

Cite this