Download presentation
Presentation is loading. Please wait.
Published byMarilynn Evans Modified over 8 years ago
1
VIPIN VIJAYAN 11/11/03 A Performance Analysis of Two Distributed Computing Abstractions
2
Outline Introduction Problem AllPairs Work Queue & Condor Evaluation Preliminary Results
3
Introduction In a biometric matching experiment, a query image is compared against a target image to compute a match score. F is the matching function. M ij = F(A i,B j ) BiBi AiAi F(A 1,B 1 ) Similarity matrix, M
4
Problem Problems encountered while computing the similarity matrix: O(n 2 ) increase in computational time where n is the number of images. Increase in computational time due to the size of the images and complexity of the matching function. Proposed solution: a distributed computing abstraction such as AllPairs.
5
AllPairs: a distributed computing abstraction As an abstraction: It simplifies the calculation of a similarity matrix: We simply have to provide the matching function, and the two sets of images. Because of the abstraction: We can take advantage of the particular abstraction to speed up computation. Ex: Taking advantage of sub-matrices.
6
AllPairs: a second look Worker 1 (Busy) Worker 2 (Busy) Worker 3 (Busy) Worker 4 (Busy) Waiting Active Here we upload 4+3 = 7 files to a single worker. As opposed to uploading 4*3 = 12 files to different workers. So the main advantage of splitting it up into sub-matrices is decrease in uploaded data.
7
Work Queue & Condor Condor: A framework which can be used to manage workload on a cluster of computers. It tries to make use of cycles from idle computers. Any Condor process can be stopped by the user of the machine. Work Queue: A system which manages Condor, which works with machines of different processing power and scheduling policies and is fault-tolerant. A stopped process, for example, is restarted automatically by Work Queue.
8
Evaluation Metrics Speed-up: Ratio of execution time using 1 processor to execution time using p processors. Scale-up: Ratio to execution time using 1 processor to execution time using p processor where problem size is scaled by p. Here we define problem size as the number of images we use. Overhead: Overall time subtracted by core computation time. This will mainly be from I/O operations and communication time between master and workers.
9
Speed-up
10
Scale-up
11
Overhead w.r.t problem size
12
Overhead w.r.t number of workers
13
Questions?
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.