Session 10: Management.

Automated Known Problem Diagnosis with Event Traces
===========================================
Idea: use SVM to classify discrete Windows system calls. 
Detail: n-grams were used as input feed. 
Comment: why not use crash stack traces? Windows already do that. Is this better? It is neither demonstrated in evaluation nor is argued.

Problem: have only tested on 4 machines. 
The classification training data size is not convincing either. 

Problem: Critised over the number of so called "root-causes" which said to have only single digit. SVM should be able to handle a larger classes with much smaller samples. not clear if SVM were really the best choice in this case.


Using Queries for Distributed Monitoring and Forensics
===========================================
Idea: Debugger for P2- causal tuples tracing, profiling...etc 
Merit: Nice for people using P2 which has no debug facility at all for the moment. Shows the power of the declarative language in distributed debugging, nice!

Problem: only useful when one is coding in P2. However, it is not that mature a general purpose language yet... even for distributed algorithms. 
Question: where is the forensics? it's distributed logging with application level causal orders.

Balancing Power Consumption in Multiprocessor Systems
===========================================
Idea: Use energy comsumption for ACPI to estimate processor temperature so as to balance the heat generation. 
Merit: Cute idea for considering heat for process migration on multi-cores. THat task energy profile is straightforward, but useful.

Problem: though heat is considered. it is not very clear what is being traded off or it can be shown that nothing is sacrificed.