Wiki

Clone wiki

gnrs / Minutes 2012-09-06

Status Report

  • The delay module will need to know the Src and Dst IP addresses in order to introduce delay.
  • Deciding to focus on a single process emulating only 1 AS's GNRS server (as opposed to multiple in a single process). This will simplify the experimental code.
  • VIA CentaurHAUL nodes supported 2100 pkts/s maximum throughput with a 1/1 (producer/consumer) thread setting. Total request delay (queue + process) increases exponentially as it approaches this limit.

Discussion

Time to Live (TTL)

  • User-predicted TTL would allow the user to estimate how long it thinks the mapping should persist.
    • For user-generated TTL, underestimation should "reward" the user, while overestimation should "punish" the user.
    • Likely related to CPU job scheduling and job length prediction.
  • Server-generated TTL would utilize information that the authoritative server has about mapping update intervals, combined with user predictions to derive its own prediction.
    • Could be predicted based on activities or user's history/expectations.
  • External services may provide additional information about appropriate TTL values.
    • For example, DHCP servers provide a lease expiration time. This can be used to generate a TTL.
  • "Soft" TTL - more of a cache hint value.
  • "Hard" TTL - a limit to how long the GUID mapping should be bound in the system.

New Tasks

Experimental

  • Let's look at the components of service time within the process: hashing time, lookup time, total time. Also evaluate kernel vs userspace time.
  • Network Delay module will use Src IP, Src Port, Dst Ip, Dst Port, Delay format. This will be the most flexible for future work and multiplexing the servers.
  • What parameters are needed for generating a GNRS workload trace? How do we obtain these parameters?

Conceptual

  • Read a few papers from previous ACM SIGMETRICS conferences.
  • List of items needed for each evaluation and the papers. What metrics are recorded, what traces to use, what topology data/ideas to evaluate, etc.
  • What to cache, how long to cache, what to replace and when to replace?
  • What will a GNRS Record include? What values, format, etc.?
  • Any prior work done for prediction/calculation of similar TTL values?

Updated