Tony's M2 Meeting Minutes
- Blesscc++ keep track of injection of all nodes: scalability solutions
- keep track of highest (how?)
- random sampling
- per region injection rate (hierarchically)
- Michael: Inline feedback more ideal
- give reflected packets dedicated virtual channel
- Onur: feeling that it has limited effectiveness compared to backchannel
- highest priorty might work
- can eliminate starvation
- must resend reflected flits
- Link-monitoring
- keep track of congestion over n>1 cycles
- at every cycle may be unfair
- very local view of network
- just looking at k is conservative, not efficient
- investigate whether k can depend on MTTS values/deflection rates
- case where every link in use but no deflections
- k depends on network location, neighbors
- k can be congestion predictor, large problem space to explore
Other Points
- duato regional congestion notification / bluegene
- router knows local and onehop deflection rates
- back network signaling: starvation response granularity
- distinguish between starvation freedom and congestion control
- target starvation freedom first
- progress to congestion control for performance, fairness, latency, bandwidth
- congestion: how to detect, when signal, who signaled
- once detected: how to throttle, when, who to throttle, how long to throttle
- feedback on throttling decisions? back back network?