Download presentation

Presentation is loading. Please wait.

Published byBrandon Amos Modified over 2 years ago

1
Author: Chengchen, Bin Liu Publisher: International Conference on Computational Science and Engineering Presenter: Yun-Yan Chang Date: 2012/04/18 1

2
Flow statistics is a basic task of passive measurement and has been widely used to characterize the state of the network. Adaptive Non-Linear Sampling (ANLS) is one of the most accurate and memory-efficient flow statistics method proposed recently. This paper studies the parameter setting problem for ANLS. Proposed a parameter self-tuning algorithm which enlarges the parameter to a equilibrium tuning point and renormalizes the counter when it overflows. 2

3
A method that adjusts the sampling rate according to the counter value. No prediction or estimation of flow size distribution is required beforehand. The counting process can be presented as c c + 1 with probability p(c) c : the counter value, 0 < a < 1. Estimate flow length can be formulated as f(c)=[(1+a) c -1]/a. Relative error of flow size can be presented as. n: real flow size 3

4
Two major performance metrics for flow statistics: Relative error Measures the accuracy of a flow statistics method and can be quantified by coefficient of variation as shown in (3) for ANLS Counting range (related to memory consumption) The largest flow size that a flow statistics method could record. For ANLS, the counting range is B=[(1+a) c -1]/a. 4

5
Tradeoff between small relative error and large counting range 5 Fig. 1. Relative error vs. parameter a when flow size n = Fig. 2. Counting range vs. parameter a when the largest counter value is 256 (the counter is 8-bit width).

6
When a counter overflows, a is adjusted from a 1 to a larger value a 2, and renormalized to a smaller value according to the reconfigured a 2. 6

7
Determine the equilibrium point for tuning the parameter. Accuracy utility (U e ) e 1 and e 2 are the relative error before and after the parameter tuning. Counting Range Utility (U b ) B 1 and B 2 are the counting range before and after the parameter tuning. 7

8
8 Fig. 3. Finding the equilibrium tuning point.

9
To keep the inverse estimations before and after the tuning is the same. Suppose the counter value and parameter before tuning is c 1 and a 1, after tuning is c 2 and a 2, the estimations are: (13) and (14) provide the same estimation 9

10
To keep the inverse estimations before and after the tuning is the same. 10 The counter is renormalized to with probability x -, and is reset to with probability 1 (x ).

11
Prove that the renormalization process does not introduce error and ensure the estimation is the same before and after renormalization. Theorem: The expect error in renormalization process is zero. Proof: Let From the Algorithm 3, the expected value of c 2 can be formulated as E(c 2 ) = (X ) + (1 X + ) = X Namely, f(E(c 2 )) = f(c 1 ). 11

12
Growth of ANLS with and without counter renormalization. 12 Fig.4 Grow of the counter value

13
Relative error of ANLS with and without parameter tuning under different traffic scenarios. 13

Similar presentations

© 2016 SlidePlayer.com Inc.

All rights reserved.

Ads by Google