Burstiness Control for High Speed Long Distance TCP

1
Burstiness Control for High Speed Long Distance
TCP

• David Wei Sanjay Hedge Steven Low
• Netlab _at_ Caltech

3rd PFLDNet Workshop Lyon, Feb 3, 2005
2
Outline

• Observations and motivation
• Burstiness Measurement
• Burstiness Control and Experimental Results
• Open problems

3
FAST over Pacific (Caltech-KAIST)

• 622Mbps, RTT360ms Caltech KAIST (Korean),
  reproduced in Dummynet alpha500 bottleneck
  buffer2000

4
Packet loss

• About 70,000 packets are lost
• But Dummynet Router dropped 30,000 packets only!

1000 sec
5
Not Only FAST! DataTAG (May 2003)

• RTT 120ms
• Bottleneck 1Gbps (split from 2.5 Gbps by Alcatel
  1670)
• Max queueing delay 2ms ( 170pkt)

6
DataTag (2003)

• DataTAG Max cwnd4000 pkt (383Mbps per flow)
• 2 flows (txqueuelen 10000)

7
Not only in Slow Start!

• DataTAG Max cwnd4000 pkt (383Mbps per flow)
• 2 flows (txqueuelen 100)

8
High Frequency Monitor

• Goal
• to understand interesting events with microscopic
  details
• low overhead (event rates 100 per ms)
• easy to use
• TCPs behavior is only determined by the
  information it receives in the last RTT.
• Implementation

9
Burstiness

• Dummynet 800Mbps, 120ms with SysKonnect NIC, 2003

10
Burstiness

• Dummynet 800Mbps, 120ms with old SysKonnect NIC

11
Packet Transmission in TCP
Packet_in_flight number of packets sent but not
acknowledged
Congestion Window Control send cwnd packets
every RTT
Sending Engine send cwnd-p_i_f packets every
acknowledgment.
12
Loop of burstiness

• Sender Ack compression/ Ack loss
• Sender A burst of data packets sent out to the
  network
• Router Packet loss due to burstiness
• Receiver Delay ack disabled more ack sent back
  to the sender
• Sender Ack compression

13
Burstiness Control?
How to Quantify?
14
Burstiness Measurement

• Long term window control guarantees that the
  average sending rate over one RTT will not exceed
  the bottleneck capacity
• Short term bursty traffic causes excessive
  queuing delay and loss
• Can we use short term excessive queuing delay
  as a measurement of burstiness?

15
Burstiness Measurement

• Green the long-term trend of queue length,
  controlled by window control algorithm
• White instantaneous queue length, the effect of
  both window control algorithm and

Queue length
time
RTT
RTT
16
Burstiness Measurement

• For each packet (packet n) to be sent, at time
  t
• u is the first packet that is sent but not
  acknowledged it is sent 1 RTT before.
• tj is the time when packet j is sent.

• Send the packet only when B is smaller than a
  threshold.

17
Burstiness Measurement

• For each packet (packet n) to be sent (u is
  the first unacked packet)

Queue length
m
u
n
time
RTT
18
Burstiness Measurement

• For each packet (packet n) to be sent (u is
  the first unacked packet)

Queue length
m
u
n
time
RTT
19
Burstiness Measurement
Available bandwidth for the connection

• For each packet (packet n) to be sent (u is
  the first unacked packet)

Queue length
m
u
n
time
RTT
20
Burstiness Measurement

• For each packet (packet n) to be sent (u is
  the first unacked packet)

Queue length
m
u
n
time
RTT
21
Burstiness Measurement
If ultm
Queue length
m
u
n
time
RTT
22
Verification of Burstiness Measurement
Red Measured Queue Length
23
Burstiness Control
24
Burstiness Control

• Max cwnd4000 pkt (383Mbps per flow)
• 2 flows

25
Burstiness Control

• 622Mbps, RTT360ms (Dummynet)

Bandwidth under-estimation by cwnd/RTT leads to
slow convergence
26
Summary

• We can control packet transmission rate in two
  levels Window control over RTT timescale and
  burstiness control over packet timescale.

Burstiness Control
Per RTT
Per Packet
Window Control
27
Open Questions

• Shall we do burstiness control in TCP?
• Increase txqueue? - only solves host problem (see
  exp)
• Simpler fixes? - scalability
• Pacing (in TCP) - overhead
• Buffer-provision in the router - expensive?
• pacing in NIC?
• Is pacing/burstiness control too polite when
  competing with unpaced TCP?
• How to reduce severe ack-compression?

28
Increasing txqueue
Txqueuelen 100 pkt
Txqueuelen 10,000 pkt
29
Thank you)

• Questions?

30
Backups
31
Continuous model

• Translate the discrete definition into a
  continuous model

32
Congestion Window Control
w/RTT
C
33
Congestion Window Control
x(t)
C
34
Congestion Window Control
x(t)
C
35
Burstiness measurement

• Theorem The queue length is bounded by the sum
  of window prediction and burstiness.

36
Summary

• We can control packet transmission rate in two
  levels Window control over RTT timescale and
  burstiness control over packet timescale.

Burstiness Control
Per RTT
Per Packet
Window Control