Bandwidth anomalous event examples

See also the library of capacity bandwidth measurements from SLAC to 40 sites for 100 days from June through September 2004. These are being used as a canonical set of data for comparing various algorithms for anomalous end-to-end network perfomance event analysis. The raw data is also available in the files starting "node" in a directory. In addition see the HW analysis of this data and the library of HW forecasts.

Also see:

SiteDatesMean changesRoute changesComments
All (40) hosts 3/5/04 - 6/14/04 16 steps, 18 diurnals, 30 hosts
Analysis of 100 days measurements with threshold set to 0.
Canonical data set 6/20/04 9/29/04 ~ 40 hosts   Raw data and Excel spreadsheets with threshold set to 0.
ANL Chicago, Illinois May 13-24 30-40% No route changes Looks like they are correlated with a job scheduled to run about 2am in the morning.
BINP Novosibirsk, Russia 6/6/04 ~50% No route change Two step changes
BINP Novosibirsk, Russia 4/23/04 46% No route change One step change
BINP Novosibirsk, Russia 9/30/04 50%, 42% No route change Two step changes
Caltech Pasadena, California 5/17/04, 5/27/04 ~35% No route change Two diurnal changes
Caltech Pasadena, California 4/8/04 - 4/11/04 ~35% No route changes Host application causes apparent bandwidth changes, spot 9 with trigbuf=20 mins
Caltech Pasadena, California 4/8/04 - 4/11/04 ~35% No route changes Host application causes apparent bandwidth changes, spot 3 with trigbuf=40 mins
CESnet Prague, Czech Republic 3/15/04 46% No route change Long term step change
Daresbury Liverpool, UK 5/8/04 & 5/9/04 12% Possible route change Two step changes close together, second one several days long
INFN Milan, Italy 4/16/04 8% No route change Tiny change (8%), not worth looking at
U Indiana Bloomington, Indiana 5/10/04 - 6/6/04 26-44% No route changes Diurnal changes
Internet2 3/25/04 85% No route change Two step changes in very bursty time series
NERSC, Oakland, California 3/21/04 28-37% No route change Step up, followed by step down lasting 1.5 days, causing 2 events
NIKHEF Amsterdam, Netherlands 5/14/04 & 5/20/04 89% Route changes Two major step changes, each causing 2 triggers, each lasting couple of days
NIKHEF Amsterdam, Netherlands 4/29/04 81% Route change Step change for about 6 hours, causing one event
NIKHEF Amsterdam, Netherlands 6/22/04 - 6/29/04 34% No Route change Step change on capacity, no route change, with threshold change of 34% & statistical limit of 1.7 standard deviations
NIIT Rawalpindi, Pakistan 4/29/04 81% Route change Lots of missing data to confuse things
ORNL Knoxville, Tennessee 4/13/04 45% No route change Outage followed by step change with 2 events
ORNL Knoxville, Tennessee 4/13/04 45% No route change Longer term view of above example
TRIUMF Vancouver, Canada 3/31/04 90% No route change Single step change, one event
U Michigan Ann Arbor, Michigan 3/25/04 58% Route changes Lots of route changes, unstable, 2-3 hour step
U Florida Gainesville, Florida 04/10/04 - 05/04/04 ~ 54-60%   Diurnal changes
U Florida Gainesville, Florida 6/22/04 - 6/29/02 No events   Illustrates history buffer mean for length of 10 hours lagging the data by many hours
U Florida Gainesville, Florida 6/22/04 - 6/29/02 No events   Illustrates history buffer mean for length of 30 hours tracking the data reasonably well
U Florida Gainesville, Florida 6/22/04 - 6/29/02 No events   Illustrates Abw history buffer mean for length of 24 hours cancelling out diurnal fluctuations
U Florida Gainesville, Florida 6/22/04 - 6/29/02 No events   Illustrates Abw history buffer mean for length of 3 days (4320 mins) which is very flat
Indiana 6/20/04 - 9/10/04 Diurnal changes plus a step change
Comparison of Plateau, K-S & -W (X^2)
Caltech 7/1/04 - 7/12/04 No alerts
Effect of smoothing paramter on synchronization of forecast and observations
Caltech and NIIT 6/19/04 - 7/31/04 No real alerts, only diurnals
Effect of missing data on H-W algorithm