Transcript Example 1
Vagueness of PlanetSeer's Measurement
Presented by Karl Deng
Example 1: Number of anomalies
“Passive monitoring allows us to detect more anomalies in less
time: we have confirmed nearly 272,000 anomalies in three
months. This is roughly 3,000 a day, and is 10 to 100 times
more than reported previously.”
What do these numbers tell us?
PlanetSeer can detect far more anomalies in less time?
Example 1: Number of anomalies
“Passive monitoring allows us to detect more anomalies in less
time: we have confirmed nearly 272,000 anomalies in three
months. This is roughly 3,000 a day, and is 10 to 100 times
more than reported previously.”
What is the number of unique anomalies among all these
anomalies?
PlanetSeer tends to duplicatedly report anormalies.
A same anomaly might be duplicately reported for a lot of
times, e.g., 1000 times.
Example 1: Number of anomalies
Without the knowledge of “to what extent anomalies are
duplicatedly reported”, we don’t get the idea of number of
unique anomalies detected by PlanetSeer.
We can only guess: Maybe PlanetSeer can detect more
anomalies than previous approaches.
Example 2: Anomaly Distributions
“Due to our wide coverage, we see new failure distribution and
location properties. … Tier 3 seems to be the most problematic,
accounting for almost half of the loops, path changes, and path
outages that we see.”
Example 2: Anomaly Distributions
Loops
Route change and outage
Number of hops in loops
Example 2: Anomaly Distributions
“Due to our wide coverage, we see new failure distribution and
location properties. … Tier 3 seems to be the most problematic,
accounting for almost half of the loops, path changes, and path
outages that we see.”
To what extent we can trust these distribution numbers?
(Distribution of Samples ≠ Real Distribution)
How good are the measurement samples?
How good are the samples?
(MonDs run on 120 CoDeeN nodes in North America)
1. Paths between CoDeeN and the clients
2. Intra-CoDeeN paths
3. Paths between CoDeeN and the origin servers.
Example 2: Anomaly Distributions
“Due to our wide coverage, we see new failure distribution and
location properties. … Tier 3 seems to be the most problematic,
accounting for almost half of the loops, path changes, and path
outages that we see.”
The sampled paths might have strong bias.
PlanetSeer might report much more anomalies in certain
regions than other regions.
Maybe Tier 3 is the most problematic.
Summary from the two examples
Vagueness 1: We have no idea on the number/distribution of
unique anomalies.
No estimation on the degree of duplicate reports is
provided.
Vagueness 2: We have no idea on how representative the
measurement results are.
No estimation on measurement errors is provided.
• Errors caused by the bias of samples.
• Errors caused by the bias of measurable samples
(due to the conservative methodology).
Example of Conservative Methodology
When analyzing the scope of path outage, the author tells us:
“Among all the outages, about 47% have no complete
reference path. In the following, we use only those with
complete reference paths in the scope analysis.”
Oh, nearly half of the samples are removed because they are
not measurable!
How much will this affect the representativeness of the
measurement results (quantitative results)?
Conclusion
Vagueness 1: We have no idea on the number/distribution of
unique anomalies.
Vagueness 2: We have no idea on how representative the
measurement results are.
The measurement results have very limited use.
Many conclusions are not reliable.