Issue

• Improve slow start to ramp up to top speed faster
• Improve congestion control algorithms to scale shrinking bandwidth more appropriately
• The current congestion control algorithm doesn't scale to IB speeds. For example:
  • You slowly climb up to get up to 10GB/s then you miss one packet and suddenly the congestion control algorithms drop you down to 5GB. The two problems are:
    • The rate of acceleration to get up to the top speed.
    • The amount of drop is excessive for those high speeds
• So that has to do with how TCP works on the inside: see slow start and tcp congestion control algorithms like bicubic and Reno.
• Neither were designed for things like IB where the difference between 0 and full speed is multiple orders of magnitude greater than in the old Ethernet/modem days.