A flexible shared library profiler for early estimation of performance gains in heterogeneous systems

Matoga, Adrian; Chaves, Ricardo; Pfister, Tomas; Roma, Nuno

doi:10.1109/hpcsim.2013.6641454

Cited by 4 publications

(1 citation statement)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Kernel profiling is a key technique used to get information about kernels to be classified, e.g. retrieving from history with profile data (Sato et al, 2011) or developing a framework for profiling a shared library (Matoga et al, 2013). A compiler framework to collect and classify kernels suitable for different devices was proposed in (Lopez-Novoa et al, 2015;Wen et al, 2014).…”

Section: Introductionmentioning

confidence: 99%

Classification of OpenCL Kernels for accelerating Java Multi-agent Simulation

Penbharkkul¹,

Marurngsith

2018

Linköping Electronic Conference Proceedings

View full text Add to dashboard Cite

Java-based multi-agent simulation (MAS) can be offloaded to graphical processing units (GPU) and other OpenCL accelerators to achieve many hundredfold speedups. However, the performance gain from the accelerated code depends strongly on whether the computation (kernels) have been scheduled to the appropriate devices. Thus, accelerating Java MAS may not lead to a sustainable speedup. This paper proposes a method for a kernel classifier to specify suitable devices to execute OpenCL kernels. The classifier can identify suitable OpenCL devices for kernels based on the static and dynamic characteristics of the code of the kernels. Kernels are grouped by their suitability for particular devices using the multiclass support virtual machine technique. After that, kernels are scheduled to an appropriate task queue. Kernel scheduling based on the proposed technique is compared against the firstcome-first-serve (FCFS) technique and against oracle scheduling when handling eight kernels. Our results show that, using the proposed method, all kernels finished execution 45 percent sooner than using the FCFS technique. However, the overall execution time was 22.5 percent longer than with oracle scheduling. Our results seem to confirm that kernel classification techniques might contribute towards sustainable high performance in accelerated Java-based MAS models.

show abstract