Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

Voodoo Constants

  •  The round-trip time
  • The total number of available priorities
  • How should we divide the prios among the folowing packet types
    • Control packets (ie. grants)
    • Unscheduled packets (priority cut off among different message sizes)
    • Scheduled packets
    • Low priority redundant preemptive scheduled packets to avoid bubbles

Homa Paper ToDos

  1. Simulation comparison points, ordered by importance

    1. self comparison when Homa features are removed

    2. pHost

    3. pFabric
    4. PIAS

    5. DCTCP

    6. investigate what causes simulation instability in phost and pfabric cods at high loads. possible explanations are uneven link utilization, more than hundred percent input rate to network, and wasted bandwidth.
  2. Complete the analysis of simulations with multiple senders and receivers:

    1. wasted bandwidth at receiver

    2. cumulative time average of priority usages

    3. time series of outstanding messages

  3. Algorithmic shortcomings

    1. if sender stops sending, should receiver "deactivate" message so it doesn't consume a priority level for redundancy?
    2. possibly receive multiple messages on the same priority level

    3. choosing parameters:

      1. dividing priorities between scheduled and unscheduled
      2. picking cutoffs for unscheduled priorities
      3. choosing redundancy factor for scheduled messages
    4. measuring actual workload, using it to pick parameters and change dynamically