Table 1: Performance modeling

1
Table 1 Performance modeling
2
Table 1 Performance modeling,Whether you
should believe my machine learning
results,Connections to static program analysis,
and Why you should drive a Lexus
3
Different Performance Models Weve Seen

• Archana Kernelized CCA
• Peter Charles Quantile regression
• Charles George Networks of queues
• QUESTION
• Do these people talk to each other?

4
Desiderata

• Confidence measures and handling non-stationarity
• Being able to give reasonable feedback to
  operator
• Like Lexuses that beep at you
• Some models can be useful without acting
• In debugging situations, cost of false positives
  is high
• Even humans have trouble in labeling data
• Also in resource allocation

5
Then Moises and Peter and I starting talking

• We agreed
• If our model worked, it would be cool
• Our model works on our test application
• We disagreed
• Peter and I We think our model will work in
  other situations
• Moises Not so much

6
Verifying Machine Learning Systems

• Two verification challenges (compared to other ML
  applications)
• Representative test sets are hard
• (Ask the people who develop bechmarks.)
• But you also need representative training sets
• e.g., very easy to model performance of my code
  Its all slow!