Congestion Control: TCP & DC-TCP

Congestion Control: TCP & DC-TCP Swarun Kumar With Slides From: Prof. Katabi, Alizadeh et al.

10Mb/s S1 R1 2Mb/s D 100Mb/s S2 S1 S2 Congestion: A big problem • Sources in the Internet are compete for bandwidth and buffer space • Why is it a problem? • Sources are unaware of current state of resource • Sources are unaware of each other • Manifestations: • Lost packets (buffer overflow at routers) • Long delays (queuing in router buffers)

10Mb/s S1 R1 2Mb/s D 100Mb/s S2 Solution: Congestion Control • Ask the senders to slow down • But how much do you slow down? - Efficiency: Maximize Utilization S1 + S2 = 2Mb/s S1 = S2 = 1Mb/s • - Fairness: Each user gets a fair share

S1 9Mb/s R1 D S2 S3 1= 1Mb/s 2= 7Mb/s 3=  Demands 1= 1Mb/s 2= 4Mb/s 3= 4Mb/s Max-min Fair Rates Max-Min Fairness • Each user gets min(demand, fair share)

TCP Congestion Control • Simple idea: Send a few packets. • If a packet is dropped => decrease rate. • If no drops => increase rate. • Control rate via congestion window • cwnd = Number of packets a sender can send without an acknowledgment (i.e. in one RTT) • Increase cwnd to increase rate • How does cwnd relate to actual rate? Throughput = cwnd/RTT

How much should TCP increase/decrease cwnd? • Basic rule: Additive increase, Multiplicative decrease (AIMD) • For every RTT: • No loss => cwnd = cwnd +1 • A loss => cwnd = cwnd/2

cwnd += 1 cwnd = 2 cwnd = 1 cwnd = 3 cwnd = 4 D A D D A A D D D A A A D D D D D AdditiveIncrease Src Dest MultiplicativeDecrease cwnd = 10 cwnd = 5 … … X

(x1/2+a,x2/2+a) (x1/2,x2/2) AIMD Leads to Efficiency and Fairness fairness Line x1=x2 (x1,x2) User 2: x2 Efficiency line (x1+x2 = C) User 1: x1

DC-TCP: TCP for Data Centers • TCP suffers major problems: • Incast • Queue Buildup • Data Center traffic is two fold: • Short Flows (50 KB – 1 MB) • Large flows (1MB – 50 MB) Delay-sensitive Throughput-sensitive

Incast Worker 1 • Synchronized mice collide. • (e.g. in map/reduce) Master Worker 2 Worker 3 Time out after 300 ms Worker 4 TCP timeout

Queue Buildup Sender 1 • Elephant flows buildup queues. • Increased latency for mice flows. Receiver Sender 2

Solution: Leverage “ECN” Sender 1 ECN = Explicit Congestion Notification ECN Mark (1 bit) Receiver Sender 2

DC-TCP: Key Idea React in proportion to the extent of congestion, not its presence.

DC-TCP Algorithm B K Don’t Mark Mark At Router: • Mark packets whenQueue Length > K. • At Sender: • Maintain running average of fractionof packets marked (α). • In each RTT: • Adaptive window decreases: • Note: decrease factor between 1 and 2.

Sample Questions (Quiz 2, S’11) Question 1: What’s the max-throughput (in pkts/s) if the source uses cwnd = 5 pkts? (Assume ACKs are delivered instantly) Capacity: 100 pkt/s Router Capacity: 10 pkt/s D S1 One-way Delay: 0.1s One-way Delay: 1s 100 pkt buffer

Sample Questions (Quiz 2, S’11) Question 2: What’s the max-throughput (in pkts/s) if the source uses cwnd = 20 pkts? (Assume ACKs are delivered instantly) Capacity: 100 pkt/s Router Capacity: 10 pkt/s D S1 One-way Delay: 0.1s One-way Delay: 1s 100 pkt buffer

Sample Questions (Quiz 2, S’11) Question 3: What’s the maximum cwnd that TCP’s AIMD can have? Capacity: 100 pkt/s Router Capacity: 10 pkt/s D S1 One-way Delay: 0.1s One-way Delay: 1s 100 pkt buffer

Sample Questions (Quiz 2, S’11) Question 4: What’s the maximum cwnd that DC-TCP can have, assuming K=40 at the router? Capacity: 100 pkt/s Router Capacity: 10 pkt/s D S1 One-way Delay: 0.1s One-way Delay: 1s 100 pkt buffer

Congestion Control: TCP & DC-TCP