WCET Measurement-based and Extreme Value Theory Characterisation of CUDA Kernels
Ref: CISTER-TR-141009 Publication Date: 8 to 10, Oct, 2014
WCET Measurement-based and Extreme Value Theory Characterisation of CUDA Kernels
Ref: CISTER-TR-141009 Publication Date: 8 to 10, Oct, 2014Abstract:
The massive computational power of graphics processor units (GPUs), combined with novel programming models such as CUDA, makes them attractive platforms for many parallel applications. This includes embedded and real-time applications, which, however, also have temporal constraints: computations must not only be correct but also completed on time. This poses a challenge because the characterisation of the worst-case temporal behaviour of parallel applications on GPUs is still an open problem. To address this situation, this paper proposes a measurement-based and statistical approach for the probabilistic characterisation of the worst-case execution time of such an application.
Kostiantyn Berezovskyi
Luca Santinelli
Konstantinos Bletsas
Eduardo TovarDocument:
View
Download22nd International Conference on Real-Time Networks and Systems (RTNS 2014).
Versailles, France.
Record Date: 8, Oct, 2014
Short links for this page: www.cister-labs.pt/docs/cister_tr_141009 www.cister-labs.pt/docs/1029