I/O Workload Overview of the Applications on Intrepid Supercomputer - - PowerPoint PPT Presentation

▶
i o workload overview of the applications on intrepid
SMART_READER_LITE
LIVE PREVIEW

I/O Workload Overview of the Applications on Intrepid Supercomputer - - PowerPoint PPT Presentation

Jun 20, 2023 637 likes •1.02k views

I/O Workload Overview of the Applications on Intrepid Supercomputer Pablo J. Pavan , Valria S. Girelli, Jean L. Bez, Francieli Z. Boito, Philippe O. A. Navaux WSPPD 2018 September 05, 2018 I/O Workload Overview of the Summary

slide-1
SLIDE 1

I/O Workload Overview of the Applications on Intrepid Supercomputer

Pablo J. Pavan , Valéria S. Girelli, Jean L. Bez, Francieli Z. Boito, Philippe O. A. Navaux

WSPPD 2018 September 05, 2018

slide-2
SLIDE 2

I/O Workload Overview of the Applications on Intrepid Supercomputer

Summary

  • Motivation
  • Goal
  • Methodology
  • Results
  • Conclusion

2

slide-3
SLIDE 3

I/O Workload Overview of the Applications on Intrepid Supercomputer

Motivation

  • HPC systems allow complex simulations

○ Medical ○ Oil and Gas Exploration ○ Weather forecasting

  • I/O is an important part of HPC applications

3 Data collected in the SantosDumond Supercomputer with the Olam application

slide-4
SLIDE 4

I/O Workload Overview of the Applications on Intrepid Supercomputer

Motivation

  • Characterizing I/O behavior can contribute:

○ I/O scheduler reconfiguration ○ I/O Stack reconfiguration ○ Creating or improving burst-buffer ○ I/O Forwarding

4

slide-5
SLIDE 5

I/O Workload Overview of the Applications on Intrepid Supercomputer

GOAL

Investigate the I/O workload of the applications executed on the Intrepid Supercomputer

5

slide-6
SLIDE 6

I/O Workload Overview of the Applications on Intrepid Supercomputer

Methodology

  • DARSHAN - HPC I/O Characterization Tool

○ Developed by Argonne L. C. F. ○ Profile I/O operations at the application level ○ Application information

■ Application identifier ■ Job identifier ■ User identifier ■ Number of MPI processes ■ Execution time

6

slide-7
SLIDE 7

I/O Workload Overview of the Applications on Intrepid Supercomputer

Methodology

  • DARSHAN - HPC I/O Characterization Tool

○ Developed by Argonne L. C. F. ○ Profile I/O operations at the application level ○ Counters

■ Individual/Collective Access ■ Interface (POSIX and MPI-IO) ■ Access sizes ■ Data Transferred ■ Number of I/O operation ■ Time spent in I/O operations

7

slide-8
SLIDE 8

I/O Workload Overview of the Applications on Intrepid Supercomputer

Methodology

  • Supercomputer Intrepid Blue Gene/P, Argonne USA

○ 91,994 captured jobs during 2012 ○ Anonymized information

8

slide-9
SLIDE 9

I/O Workload Overview of the Applications on Intrepid Supercomputer

Methodology

  • Our process to analyze the DARSHAN data

9

slide-10
SLIDE 10

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Overview of the 2012 dataset

10

slide-11
SLIDE 11

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Overview of the 2012 dataset
  • The analysis of the applications whose I/O percentage

was above 50%

11

slide-12
SLIDE 12

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Overview of the 2012 dataset
  • The analysis of the applications whose I/O percentage

was above 50%

  • Top ten most executed applications during 2012

12

slide-13
SLIDE 13

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Overview of the 2012 dataset

○ 91,994 jobs resulted in 26,034 different applications ○ Executions with a high I/O percentage do not run for long periods

  • f time

■ 45,000 seconds executing ■ 2,500 seconds performing I/O

13

slide-14
SLIDE 14

I/O Workload Overview of the Applications on Intrepid Supercomputer

  • 85,710 jobs
  • 10,050 jobs have an I/O

percentage above 50%

Results

14

  • Overview of the 2012 dataset

○ 91,994 jobs resulted in 26,034 different applications ○ Executions with a high I/O percentage do not run for long periods

  • f time

■ 45,000 seconds executing ■ 2,500 seconds performing I/O

slide-15
SLIDE 15

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • The analysis of the applications whose I/O percentage

was above 50%

○ We selected three applications

15

Exec (Anonymised) Jobs with > 50%

  • f the time in

I/O operations 1176110786 980 1338247359 898 902685977 761 Total 2.639

slide-16
SLIDE 16

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • The analysis of the applications whose I/O percentage

was above 50%

  • Exec 1176110786

○ 1,024 or 2048 MPI processes ○ Executed 980 times ○ From June 06, 2012 to December 18, 2012

16

slide-17
SLIDE 17

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • The analysis of the applications whose I/O percentage

was above 50%

  • Exec 1176110786

○ 1,024 or 2048 MPI processes ○ Executed 980 times ○ From June 06, 2012 to December 18, 2012

17

slide-18
SLIDE 18

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • The analysis of the applications whose I/O percentage

was above 50%

  • Exec 1176110786

○ 1,024 or 2048 MPI processes ○ Executed 980 times ○ From June 06, 2012 to December 18, 2012

18

slide-19
SLIDE 19

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • The analysis of the applications whose I/O percentage

was above 50%

  • Exec 902685977

○ 32, 64 and 128 MPI processes ○ Executed 763 times ○ From October 23, 2012 to October 30, 2012

19

slide-20
SLIDE 20

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

20

slide-21
SLIDE 21

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

21

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-22
SLIDE 22

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

22

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-23
SLIDE 23

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

23

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-24
SLIDE 24

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

24

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-25
SLIDE 25

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

25

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-26
SLIDE 26

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

26

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-27
SLIDE 27

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

27

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-28
SLIDE 28

I/O Workload Overview of the Applications on Intrepid Supercomputer

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

28

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

slide-29
SLIDE 29

I/O Workload Overview of the Applications on Intrepid Supercomputer

Exec (Anonymised)

Number of Observations

Execution time (seconds) I/O Time (seconds) I/O % Data transferred (GB) 685531913

8.016

1.370 3,12 0,22 0,003 1330277471

5.176

1.314 536,63 40,84 5,69 931947437

4.922

2.906 985,65 33,91 0,013 3069475893

3.191

2.808 3,92 0,13 0,012 1074553177

2.690

1.489 533,55 35,83 1,12 1648769576

2.588

1.993 191,82 9,62 7,59 1633035531

2.259

42.003 8.176,28 19,46 6223,78 2425255765

1.339

3.720 2.967,61 79,77 2,85 3475271559

1.303

1.512 310,14 20,51 2,25 1338247359

1.051

2.415 1.924,92 79,70 17,80

Results

  • Top ten most executed applications during the year
  • The total executions count was 32,535
  • Which represents 35.36%

29

This case can indicate:

  • Problem with the program level
  • I/O contention
slide-30
SLIDE 30

I/O Workload Overview of the Applications on Intrepid Supercomputer

Conclusion

  • We analyzed the results of I/O workload found during

the year 2012 on the Intrepid supercomputer

  • We obtained 91,973 submitted jobs that represent

26,034 different applications

30

slide-31
SLIDE 31

I/O Workload Overview of the Applications on Intrepid Supercomputer

Conclusion

  • Applications that had spend more than 50% of their

time in I/O

  • We could notice different behaviors:

○ MPI process numbers ○ Transferred data ○ Execution period during the year

31

slide-32
SLIDE 32

I/O Workload Overview of the Applications on Intrepid Supercomputer

Conclusion

  • Applications that had more than 50% of spend time I/O
  • We could notice different behaviors:

○ MPI process numbers ○ Transferred data ○ Execution period during the year

  • Top ten most executed applications

○ Representing 35.36% of the total number of executions observed ○ In seven applications: ■ More than 19% of their execution time spent in I/O time

32

slide-33
SLIDE 33

I/O Workload Overview of the Applications on Intrepid Supercomputer

Conclusion

  • As future work

○ We intend to extend this analysis of the characterization of the intervals for all the jobs

  • f Intrepid seeking to group applications that

have a similar I/O behavior

33

slide-34
SLIDE 34

I/O Workload Overview of the Applications on Intrepid Supercomputer

Conclusion

  • As future work

○ We intend to extend this analysis of the characterization of the intervals for all the jobs

  • f Intrepid seeking to group applications that

have a similar I/O behavior

  • We also intend to aggregate data from multiple

applications to have a global view of the use of I/O in a supercomputer

34

slide-35
SLIDE 35

I/O Workload Overview of the Applications on Intrepid Supercomputer

Acknowledgment

  • This research received funding from the Petrobras

project, grant n. 2016/00133-9. It was also supported by PIBIC CNPq-UFRGS and PROBIC FAPERGS-UFRGS

  • This research used resources of the Argonne

Leadership Computing Facility at Argonne National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under contract DE-AC02-06CH11357

35

slide-36
SLIDE 36

I/O Workload Overview of the Applications on Intrepid Supercomputer

Pablo J. Pavan , Valéria S. Girelli, Jean L. Bez, Francieli Z. Boito, Philippe O. A. Navaux

WSPPD 2018 September 05, 2018