Presentation is loading. Please wait.

Presentation is loading. Please wait.

WCET Measurement-based and Extreme Value Theory Characterisation of CUDA Kernels Kostiantyn Berezovskyi CISTER ISEP/IPP Porto, Portugal Konstantinos Bletsas.

Similar presentations


Presentation on theme: "WCET Measurement-based and Extreme Value Theory Characterisation of CUDA Kernels Kostiantyn Berezovskyi CISTER ISEP/IPP Porto, Portugal Konstantinos Bletsas."— Presentation transcript:

1 WCET Measurement-based and Extreme Value Theory Characterisation of CUDA Kernels Kostiantyn Berezovskyi CISTER ISEP/IPP Porto, Portugal Konstantinos Bletsas CISTER ISEP/IPP Porto, Portugal Eduardo Tovar CISTER ISEP/IPP Porto, Portugal Luca Santinelli Onera Toulouse, France Supported by National Funds through FCT (Portuguese Foun dation for Science and Technology) and by ERDF (European Regional Development Fund) through COMPETE (Operational Programme ’Thematic Factors of Competitiveness’), within projects ref. FCOMP FEDER (CIS- TER) and FCOMP FEDER (REGAIN); by FCT and the EU ARTEMIS JU funding, within project ARTEMIS/0001/2013, JU grant nr (EMC2); by FCT and by ESF (European Social Fund) through POPH (Portuguese Human Potential Operational Program), under PhD grant SFRH/BD/82069/2011.

2 Stream Processing Kernels Series of operations. Input: streams. Output: streams. 2 Streams Collection of data. All data is expressed in streams.

3 Why Streams? Data parallelism Stream elements can be processed at once. Task parallelism Pipeline. © nvidia.com 3

4 GPU software application Minimal dependency between data elements. 4 Large data collections. Data parallelism. High arithmetic intensity.

5 Application areas © nvidia.com 5

6 © Kirk, David B. and Hwu, Wen-mei W. GPU as a co-processor

7 7 Related work

8

9 9 16 streaming multiprocessors GPU Architectures © nvidia.com Fermi Kepler 15 streaming multiprocessors

10 Previous work Exact Upper bound on the

11 Previous work

12 Probabilistic WCET (pWCET)

13 Estimating pWCET MeasurementsStatistical approach How to make it safe?

14 Measuring kernel timings © nvidia.com

15 Measuring kernel timings

16

17

18 Hardware architecture © nvidia.com

19 Hardware architecture © nvidia.com

20 Hardware architecture © nvidia.com

21 Streaming multiprocessor

22 Measurement technique as © nvidia.com

23 Measurement technique © nvidia.com

24 Measurement technique

25 © oracle.com

26 Measurement technique © oracle.com© nvidia.com

27 Measurement technique

28 The case-study kernel GeForce GTX 770 Graphics Card 8 streaming multiprocessors Thread Block -- the group of threads that are processed by a single streaming multiprocessor.

29 Extreme value theory (EVT) Statistical approach Block MaximaPeaks-Over-Threshold

30 Independence, stationarity and extremal tools

31 Single thread block experiments

32 32 thread blocks experiments

33 Measurement extremogram

34 EVT estimates

35 CDF representation of the distributions

36

37 Related work

38 Future Work

39 39 Questions?

40 40 Thank You!


Download ppt "WCET Measurement-based and Extreme Value Theory Characterisation of CUDA Kernels Kostiantyn Berezovskyi CISTER ISEP/IPP Porto, Portugal Konstantinos Bletsas."

Similar presentations


Ads by Google