Network
Weather Forecasting
Domain:
The analysis of data from High Energy Physics is increasingly dependant on high speed networks to share the data across world-wide collaborations. High speed networks often suffer with sudden drops in performance due to various reasons. The current approach to network measurement and diagnosis is generally reactive by manually viewing graphs and tables of analyzed measures and results. We assume that an automated approach to forecast and event detection can serve network engineers and administrators better to identify network problems.
Problem:
To forecast values of network performance into future with reasonable confidence so that preventive measures can be adapted before hand.
Related work:
Significant work is being done for two some what related projects i.e., Network anomaly detection and Network event isolation. In Network anomaly detection the approach is to find out unwanted events in networks through rigorous analysis of data collected using different measuring tools. Event diagnosis is an extension to Network anomaly detection and goal is to classify unwanted events on the basis of their cause, effect and other differences.
Network weather forecasting is similar to above two projects in a sense that it applies similar rigorous statistical analysis and can be used for anomaly detection but it is different on a ground that it is proactive and both other approaches are reactive so in a way it can prove more efficient than other.
Previous work:
Some work has
already been carried out for network weather forecasting. That work includes
forecasting of active and passive data and its comparison. Main algorithm
tested for this purpose was Holt-Winters. Although the results are encouraging,
yet we believe some more sophisticated scheme is required to improve the
results.
Our Goals
1- We have to get a set of data that can be used as basic data set for testing.
2- We have to apply regularization where needed.
3- Implementation of main stream ARMA/ARIMA forecasting algorithm
4- Apply algorithm on previous dataset
5- Apply algorithm on new dataset (if any)
6- Results analysis
Task division and dependency
Following table shows a rough distribution of tasks identified in this project, their dependence on other tasks.
PX = PhaseX
TX = TaskX
|
Phase |
Task |
Description |
Dependence |
|
P1 |
T1 |
Previous
work study |
Nil |
|
P1 |
T2 |
Study
of ARMA/ARIMA |
Nil |
|
P1 |
T3 |
Getting
previously tested data set/results |
Nil |
|
P2 |
T4 |
Design
of software |
T2 |
|
P2 |
T5 |
Implementation |
T4 |
|
P3 |
T6 |
Regularization
|
T3,T5 |
|
P4 |
T7 |
Apply
on previously tested data set |
T3,T5,T6 |
|
P4 |
T8 |
Analysis/Comparison
with previous results |
T7,
T3 |
|
P5 |
T9 |
Apply
on any new dataset |
T5,T6 |
|
P5 |
T10 |
Analysis |
T9 |
|
P6 |
T11 |
Report
Results |
T8,T10 |