• An extensible OS permits modification of kernel behavior
• How do we know which parts of the system to extend?
  Answer: Self-monitoring.
• By what mechanism is the system actually extended?
  Answer: Self-adaptation.

• VINO is an extensible system that can be made to do self-monitoring and self-adaptation.
• Extensibility is provided using grafts.
• Grafts allow the following types of extensibility:
  1. Replacing a method of an object.
  2. Registering a handler for a kernel event.

1. Continuous monitoring of the system to construct a database of performance statistics.
2. Correlation of the database by process, process type, and process group.
3. Collection of traces and logs of process activity.
4. Deriving heuristics and algorithms to improve performance for the observed patterns.
5. In Situ simulation of new algorithms using statistics, logs and traces.
6. Adapting the system according to the results of simulation.

• Each VINO subsystem includes a statistics module that maintains counts of events handled by the subsystem.
• A graft is created to poll each of these subsystems at a regular interval.
• SUIF compiler (Harvard) can also be used to allow useful profiling.

• Capture dynamic data about the system.
• A trace is a record of requests coming into a module.
• A log is a record of requests going out from a module.

• Statistics, traces, and logs create a picture of what is happening on a system.
• VINO allows a module under investigation to be replaced, rerun, and the results compared.
• Simulations run along side of other processes on the system (in situ) but do not affect the state of the system.

• There are two types of analysis mechanisms, off-line and on-line.
• Off-line analysis is performed at night and is responsible for the following:
  1. How frequently measurements should be taken.
  2. Setting thresholds for each resource.
  3. Long-term performance of the system.
• On-line analysis is performed continuously and is responsible for the following:
  1. Determining instantaneous resource utilization and rate of change.
  2. Detects "red-flag" situations as presented by the off-line system.
  3. Selects adaptation heuristics or installs the appropriate trace-generating graft.

• Paper mentions that precisely how the system adapts itself is one of the key areas for future research.
• A few simple adaptation techniques are presented at the end of the paper.

Problems with the Paper

• Good techniques in theory: do they really work???
  How much overhead required?
  No Performance Measurements data at all.
• Very little discussion on alternative methods/grafts
  • Source
  • Security
  • Number of Grafts
  • Performance Overhead
• Discusses only the simple cases (i.g. Linear Paging)