Status Report

• Previous results showing that request execution time decreasing with increasing arrival rate was shown to be a side-effect of frequency scaling.
  • By fixing the CPU frequency, we now see that the average execution time for servicing a request increases as the arrival rate increases.
  • This is not fully explained by looking at the mean, however, so we will investigate the cause for next week.
• Work on the delay module is being planned, and two options are available for loading the delay values. The first is loading via message-passing (NetLink sockets), the other is making use of the Click! proc filesystem for element configuration (/click/<element>/config).
  • Decision was made to use the file system, since this will enable the most rapid development for this stage.

Discussion

Caching of GUID Mapping Records

• If no caching of the GUID -> NA mapping is performed, every look-up requires a large network traversal. So where to cache and what should be cached?
  • One option is to have a Local Name Resolution Service (LRNS) that caches all requests and inserts known to the AS.
  • Another option is routing "along the way", but would there be a measurable benefit to this approach?
• Without knowing an expected workload, how can we design a caching protocol?
  • We can use a trace of HTTP traffic to estimate a back-of-the-envelope caclulation.
• Of note - the testbed does not support caching, since it is using point-to-point communication.
• Another idea is allowing the network to generate statistics of how frequently the updates occur.
• Two tracks to handle caching
  • Start an implementation of a caching server by ripping-out some functionality from the main server
  • Start to think about workloads, interfaces, caching policies.

Expiration/Ejection of GUID Mapping Records

• Can we use the TTL field to hint or force cache ejection?
• Should mappings support an expiration field, or should they be expired explicitly?
• How do we handled orphaned mappings. That is, mappings that are no longer valid, but which will never be updated/expired/managed again?

New Tasks

• Investigate the cause of execution time increase under increasing request rate.
• Allow the delay module to be configured at run-time via the /click file system.
• Implement a simple caching mechanism in the GNRS server to see how the system performs with and without caching.
• Think about how we will handle content- and context-based GUID mapping.
  • What type of content will be mapped and how?
  • How big are the mapping records?
  • What is a reasonable working set for an AS's cache? How much content exists?