• Concrete metrics for congestion detection
  • count deflections at each node
  • node cannot inject: starvation
    • 4 inputs: no more than 3 = fine, below saturation
    • whenever router detects >3, sends signal
    • know sources at router
      • simple solution: send message to every source (maybe only if it needs to inject)
      • only routers that need to inject complain
  • possibly detect how many nodes are sending back reduction messages to determine how much to reduce injection
  • back channel that signals router ID to reduce injection
    • determine ID:
      1. look at any/all flit(s)
      2. determine rates coming from different sources
    • evaluate both to see if any benefit to added overhead needed for logic in b)

• looking for simple basline algorithm
• can we prove something about system/algorithms based on evaluations? (prevents starvation in some timeline, etc)

• without congestion control, no way to provide latency guarantees
  • therefore latency guarantees are a product of good/fair injection control
  • add MTTS counters to help prioritize/help latency guarantee
    
• Feedback mechanism
  • simple case is backchannel

• Processor Simulator Requirements
  • application traces
  • local cache
  • inst window
  • Operation:
    1. read from trace file: mem vs non-mem
    2. local vs nonlocal cache
    3. insert into network
    4. receive request back
  • model interconnection network as a distributed cache
  • assume everything hits within cache - a miss goes out to mem controller, not part of interconn network

Back to Project Page