Understanding the Performance of TCP Pacing

1
Understanding the Performance of TCP Pacing

• Amit Aggarwal, Stefan Savage, Thomas Anderson
• Department of Computer Science and Engineering
• University of Washington

2
TCP Overview

• TCP is a sliding window-based algorithm.
• Ack-clocking.
• Slow-start phase (W2W each RTT).
• Congestion-avoidance phase (W each RTT).

TCP Burstiness

• Slow Start
• Losses
• Ack compression
• Multiplexing

3
Motivation

• From queuing theory, we know that bursty traffic
  produces
• Higher queuing delays.
• More packet losses.
• Lower throughput.

Random
Worst Case
Best Case
Response Time
Load
1
Queue Capacity
4
Contribution

• Evaluate the impact of evenly pacing TCP packets
  across a round-trip time.

What to expect from pacing TCP packets?

• Better for flows
• Since packets are less likely to be dropped if
  they are not clumped together.
• Better for the network
• Since competing flows will see less queuing delay
  and burst losses.

5
Simulation Setup
4x Mbps 5 ms
R1
S1
x Mbps
B
40 ms
S pkts
4x Mbps 5 ms
Rn
Sn

• Jains fairness index f
• f (? xi)2 (? xi RTTi)2
• n ? xi2 n ? (xi RTTi)2

6
Experimental Results

• A) Single Flow
• case S0.25BRTT
• TCP Reno due to its burstiness in slow-start,
• incurs a loss when W0.5BRTT
• paced TCP incurs its first loss after it
  saturates the pipe, I.e when W2BRTT
• As a result, TCP Reno takes more time in
  congestion avoidance to ramp up to BRTT
• (paced TCP achieves better throughput only at
  the beginning)
• case S?BRTT
• (They both achieve similar throughput)
• The bursty behavior of TCP Reno is absorbed
  by the buffer and it does not get a loss until
  WBRTT

7

• B) Multiple Flows
• 50 flows starting at the same time. All flows
  have same RTT.
• case S0.25BRTT
• (TCP Reno achieves better throughput at the
  beginning!)
• (Paced TCP achieves better throughput in
  steady-state!)
• TCP Reno Flows send bursts of packets in
  clusters some drop early and backoff, allowing
  the others to ramp up.
• paced TCP All the flows first saturate the
  pipe. At this point everyone drops because of
  congestion and mixing of flows, thereby leaving
  the bottleneck under-utilized. (Synchronization
  effect)
• In steady state, all packets are spread out and
  flows are mixed as a result there is a
  randomness in the way packets are dropped. During
  a certain phase, some flows might get multiple
  losses, while others might get away without any.
  (De-synchronization effect)
• case S?BD
• De-synchronization effect of Paced TCP persists.
  

8

• C) Multiple Flows - Variable RTT
• 50 flows starting at the same time. 25 flows
  have RTT100 msec
• and 25 flows with RTT280 msec.
• case S0.25BRTT
• (Paced TCP achieves better fairness without
  sacrificing throughput)
• TCP Reno the higher burstiness as a result of
  overlap of packet clusters from different flows
  becomes visible. It has a higher drop rate at the
  bottleneck link while achieving similar
  throughput.
• case S?BD
• TCP Reno higher drop rate persists.

9

• D) Variable Length Flows
• A constant size flow is established between each
  of 20 senders and corresponding 20 receivers. As
  a particular flow finishes, a new flow is
  established between the same nodes after an
  exponential think time of mean 1 sec.
• Ideal-latency the latency of a flow that does
  slow start until it reaches its fair share of the
  bandwidth and then continues with a constant
  window. (just for comparison reasons)
• Phase1 no losses. Latency of paced TCP slightly
  higher due to pacing.
• Phase 2 S0.25BRTT TCP Reno experience more
  losses in slow start some flows timeout. Case
  S?BD this effect disappears.
• Phase 3 Synchronization effect of paced TCP is
  visible.
• Phase 4 Synchronization effect disappears
  because flows are so large that new flows start
  infrequently.

10

• E) Interaction of Paced and non-paced flows
• A paced flow is very likely to experience loss
  as a result of one of its packets landing in a
  burst from a Reno flow.
• Reno flows are less likely to be affected by
  bursts from other flows.
• Result TCP Reno have much better latency than
  paced flows, when both are competing for
  bandwidth in a mixed flow environment.
• If we continuously instantiate new flows, the
  performance of paced TCP even deteriorates more.
  New flows in slow start, cause the old paced
  flows to regularly drop packets, further
  diminishing the performance of pacing.

11
Conclusion
Pacing improves fairness and drop
rates. Pacing offers better performance with
limited buffering. In other cases pacing leads
to performance degradation due to 1. Pacing
delays the congestion signals to a point where
the network is already over subscribed. 2. Due
to mixing of traffic pacing synchronizes drops.