hironaka / bufferbloat Goto Github PK

Instructions: sudo ./run.sh

Max Queue Size: 20 packets Mean Transfer Time: 3.04112 Standard Deviation: 0.470129541297

Max Queue Size: 100 packets Mean Transfer Time: 7.64675 Standard Deviation: 0.98010557467

TCP congestion control algorithm attempts to achieve full utilization of the links by increasing the size of the congestion window until a packet is dropped, i.e. the output buffer at the bottleneck link is full. In our case, the long lived tcp connection will fill the buffer at the switch, regardless of its size. When a new TCP connection is created to fetch a webpage, its tcp packets are placed at the end of this queue. The longer the queue, the longer each packet must wait to be routed to the next hop, and the longer the webpage fetch takes.

1000 packet max queue size 1500 byte MTU 100 Mb/s drains

1500 * 1000 * 8 / (100 * 10^6) = 0.12 seconds

RTT ~= 6 * QUEUE_SIZE + C C = Propagation Delay + Packetization Delay ~= 20ms

One way to mitigate buffer bloat would be to differentiate TCP flows based on the Differentiated Services flags in the TCP header. This would only mitigate the problem by using multiple smaller buffers.

Another way to mitigate the problem is to employ active queue management algorithms. These algorithms intelligently choose to preemptively drop or mark packets prior to the queue filling up, preventing a long lived TCP flow from blocking other flows.