Skip to content

Latest commit

 

History

History
31 lines (28 loc) · 1.7 KB

design.md

File metadata and controls

31 lines (28 loc) · 1.7 KB
Raw

Simulator Design Summary

The simulator is event-driven (single threaded) and operates at the level of a large datacenter network topology. It takes the following inputs which are specified through a JSON input file:

  1. Topology: the topology type (e.g. FB Fabric) along with its parameters.
  2. Link failure trace: a trace containing link failure events where each event is denoted as: <time>, <link id>, <loss rate>. For example: 349200,6136,6.5e-05 denotes that at time 349200, link ID 6136 in the network topology started corrupting packets with a loss rate of 6.5e-05.
  3. Solution: The solution could either be CorrOpt which is an algorithm to disable a subset of the failed links or it could be the joint strategy of LinkGuardian + CorrOpt as proposed in our paper (section 3.6). Any parameters corresponding to the solution are also required as the input; most importantly, the "capacity constraint" as per which the solution needs to operate.

The simulator then outputs a timeseries of several topology-level performance parameters, most important of which are the following:

  1. Total penalty: sum of the loss rates for all the active (remaining) corrupting links in the network.
  2. Least paths per ToR: the least fraction of paths to the spine (top) layer of the network for the worst-case ToR. This metric captures the impact on per-ToR path diversity as corrupting links are disabled for repair.
  3. Least capacity per pod: the total capacity in a network pod from the ToR-layer to the spine (top) layer for the worst-case pod in the network.