Presentation is loading. Please wait.

Presentation is loading. Please wait.

1 IDGF International Desktop Grid Federation ASSESSING THE PERFORMANCE OF DESKTOP GRID APPLICATIONS A. Afanasiev, N. Khrapov, and M. Posypkin DEGISCO is.

Similar presentations


Presentation on theme: "1 IDGF International Desktop Grid Federation ASSESSING THE PERFORMANCE OF DESKTOP GRID APPLICATIONS A. Afanasiev, N. Khrapov, and M. Posypkin DEGISCO is."— Presentation transcript:

1 1 IDGF International Desktop Grid Federation ASSESSING THE PERFORMANCE OF DESKTOP GRID APPLICATIONS A. Afanasiev, N. Khrapov, and M. Posypkin DEGISCO is supported by the FP7 Capacities Programme under contract nr RI-261561.

2 2 Desktop Grids in a Few Words “Desktop Grids, Desktop Clouds, allow to employ otherwise idle computing time of Desktop computers for large computational programmes. Desktop Grids can be used inside an organisation, or they can collect computing time from volunteers all over the country, or even all over the world.” (IDGF Flyer) “Desktop Grids, Desktop Clouds, allow to employ otherwise idle computing time of Desktop computers for large computational programmes. Desktop Grids can be used inside an organisation, or they can collect computing time from volunteers all over the country, or even all over the world.” (IDGF Flyer)

3 3 DEGISCO WP4 10/03/2016 3 Why Understanding the Performance is Important?  Desktop Grids collect huge number (millions) of CPUs  But there are huge overheads: – nodes are not available all the time – only a part of CPU power is available – high percentage of faults – heterogeneity  The real performance is much less than the peak one

4 4 DEGISCO WP4 10/03/2016 4 Why Understanding the Performance is Important?  Know what you really gain from parallelization  Identify sources of overhead  Compare different load distribution policies and select the best

5 5 DEGISCO WP4 10/03/2016 5 The “Parallel” Speedup The speedup is defined as the ratio of the time taken to solve a problem on a single processing element to the time required to solve the same problem on a parallel computer with p identical processing elements. ( Ananth Grama, Anshul Gupta, George Karypis, Vipin Kumar, Introduction to Parallel Computing, Addison-Wesley, 2003. )

6 6 DEGISCO WP4 10/03/2016 6 The “Parallel” Speedup Doesn’t work for desktop-grid systems because of two reasons: 1.Desktop Grid applications are long running (weeks, months) => T s is too large to obtain in a reasonable time 2.The processors are NOT identical => which processor should be used as a reference?

7 7 Desktop Grid Terminology Workunits Results

8 8 Speedup Notion for Desktop Grids Speedup Total time (sum for all successful workunits) For parallel systems and thus Makespan

9 9 The Speedup as a Function of Time Interval The introduced notion of speedup requires the application to terminate – not very practical for long running tasks. We can define the speedup for a given time interval: a total time of a useful job performed between moments t 1 and t 2.

10 10 How to compute T tot (t 1,t 2 )? internal jobs external jobs boundary jobs accounted not accounted Boundary jobs are not accounted but contribute to the useful job!

11 11 How to compute T tot (t 1,t 2 )? For accurate measurement we need:

12 12 Replication Workunits Results

13 13 What to do with Replication? Replication is used in Desktop Grids to guarantee the result correctness and to achieve better load balancing We take average time over all successful replicas

14 14 Implementation Logger Web-based UI Log files BOINC periodic task Based on Google Web Toolkit

15 15 Implementation

16 16 DEGISCO WP4 10/03/2016 16 Application to OPTIMA@home  The performance evaluation tools have been deployed at OPTIMA@home project (available through administrative Web-interface)  Helped to measure the real speedup and to compare the efficiency of different load distribution strategies

17 17 Other Tools: SZTAKI Package Good tool shipped with the SZTAKI Desktop Grid: shows the performance of your project in GFlops

18 18 DEGISCO WP4 10/03/2016 18 Other Tools: Statistic Sites BOINCStats calculates the average performance

19 19 DEGISCO WP4 10/03/2016 19 The Comparison  Speedup is more informative than GFlops performance  Speedup is application oriented while the performance is infrastructure-oriented  The user should have a possibility to tune the time interval  It is good to know both speedup and performance

20 20 DEGISCO WP4 10/03/2016 20 Conclusions and Future Work  Extend the notion of parallel efficiency in a similar way  Improve web-interface  Make our tools publically available open source  Maybe integrate with some existing package e.g. SZTAKI desktop grid

21 21 http://desktopgridfederation.org/ Globe adapted from http://upload.wikimedia.org/ wikipedia/commons/f/fa/ Globe.svg IDGF International Desktop Grid Federation


Download ppt "1 IDGF International Desktop Grid Federation ASSESSING THE PERFORMANCE OF DESKTOP GRID APPLICATIONS A. Afanasiev, N. Khrapov, and M. Posypkin DEGISCO is."

Similar presentations


Ads by Google