Structure from Failure

Paper Figures

Here you can find the full results for all the figures in an earlier NIPS submission. They are taken from the more detailed list of results further down below.

• Figure 2 (a), Figure 2 (b)
• Figure 3 (a), Figure 3(b)
• Figure 4 (a), Figure 4 (b)

Synthetic Data

Here you can find the results from synthetic data. The trees were generated at random with a random parent relationship of probability of 50%.

Correct Prior

For generating the dataset, we fitted the 10%-quantile and 90%-quantile of a Gamma distribution to fit the following waiting times:

• No parent: 7 days - 25 days
• Parent up: 10 days - 30 days
• Parent down: 5 minutes - 30 minutes
• Parent unknown: 7 days - 25 days

For inferring the structure, we used the same prior distribution than for generating the data. The results are obtained by generating data for several periods of times:

• 3 months (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 6 months (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 1 year (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 2 years (dataset, ground truth, MAP, DOWS tree (for details see here), parameter match: no parent, parent up, parent down)

Vague Prior

For generating the dataset, we fitted the 10%-quantile and 90%-quantile of a Gamma distribution to fit the following waiting times:

• No parent: 7 days - 25 days
• Parent up: 10 days - 30 days
• Parent down: 5 minutes - 30 minutes
• Parent unknown: 7 days - 25 days

For inferring the structure, we fitted the 10%-quantile and 90%-quantile of a Gamma distribution to fit the following waiting times:

• No parent: 3 days - 40 days
• Parent up: 5 days - 60 days
• Parent down: 3 minutes - 2 hours
• Parent unknown: 3 days - 40 days

The results are obtained by generating data for several periods of times:

• 3 months (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 6 months (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 1 year (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 2 years (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)

Correct Prior (unreasonable settings)

For generating the dataset, we fitted the 10%-quantile and 90%-quantile of a Gamma distribution to fit the following waiting times:

• No parent: 2 days - 8 days
• Parent up: 3 days - 7 days
• Parent down: 12 hours - 1 day 12 hours
• Parent unknown: 2 days - 8 days

For inferring the structure, we used the same prior distribution than for generating the data. The results are obtained by generating data for several periods of times:

• 3 months (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 6 months (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 1 year (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)
• 2 years (dataset, ground truth, MAP, parameter match: no parent, parent up, parent down)

Server Farm of a Major Microsoft Web Site

Here you can find the results from our analysis of a server farm of a major Microsoft web site. The analysis is based on event logs from the past 5 years. For legal reasons, we have encrypted the server names. However, it is ensured that each SQL, IIS and WEB server is singled out by having [A], [B] and [C] in the encrypted server name, respectively. Our priors have been set using the 10%- and 90%-quantile of the Gamma distribution matching with domain expert knowledge of the runtimes. The DOWS results are obtained using a state-of-the-art rule based system to detect failure dependencies in large-scale networks based on the following paper B. Levidow and B. Murphy. Windows 2000 Dependability. Note that DOWS is a system that tries to detect clusters of servers which are best represented by a graph.

• CTBN (1:1,000,000 favour for root, 1:1,000,000,000 favour for root)
• DOWS (trees, graph)

Machine Learning and Perception—Machine Learning—Continuous Time Bayesian Networks—Structure from Failure

This site was last updated 17-02-2005