Title: Energy Consumption of TCP Reno, Newreno, and SACK in Multi-Hop Wireless Network
1Analysis of Energy Consumption of TCP Reno,
Newreno, and SACK in Multi-Hop Wireless Network
Thesis Defense of Harkirat Singh
2Outline
- Motivation
- TCP Mechanism
- Energy Model
- Experiment Methodology
- Experiment Results
- Conclusion
3Motivation
- Wireless devices need to be
- Reliable
- Light Weight
- Wireless links are error prone, low bandwidth
etc. -
- TCP connection suffers long idle time hence low
throughput - How is Power consumed in a wireless node?
4Lucent 802.11 Silver (11 Mbps) WaveLAN DSSS PC
Card Characteristics INFOCOM01
- Radio States Transmit, receive, idle, sleep
Power (Idle) 780mW
Power (Sleep) 50mW
Idle / Sleep
Ack
Data
Recv
Xmt
Power (Xmt) 1478mW
Power (Recv) 900mW
5TCP Mechanism
- TCP is a sliding window protocol with built in
self-clocking and congestion control algorithm - Slow start
- Congestion avoidance
- Fast retransmit
- Fast recovery
- TCP Tahoe implements first three
- TCP Reno, Newreno and SACK have all four
6Slow start
- Slow start is used during initial phase of a
connection or after retransmission timeout
(Probing the network) - It is also known as additive increase and
multiplicative decrease -
-
-
-
cwnd1 cwnd2 cwnd4 cwnd8
Sender Receiver
7Fast retransmit and recovery
- TCP receiver generates ack for next expected
packet - In the event of packet loss or reordering,
receiver Acks last in-sequence pkt received - TCP sender retransmits first unacknowledged
packet after 3 dup_acks without waiting for
retransmit timer to go off - ssthresh min (cwnd, rcv advertised window) / 2
- Send missing pkt and set cwnd 3
- Inflate cwnd till retransmitted pkt is acked
- Deflate the window, enter Congestion Avoidance
8Congestion Avoidance
- When cwnd gt ssthresh TCP invokes Congestion
Avoidance - cwnd 1/cwnd (for each ack)
- In other words cwnd is incremented (cwnd) by
one per round trip time - It is conservative approach to increase cwnd as
name suggests - When congestion is detected (by timeout), TCP
slows down and invokes slow start
9Evolution of TCP window
Fast retransmit and recovery
Slow start
Congestion avoidance
Congestion Window
Time
10Evolution of cwnd
Outstanding Data
Time
11TCP Newreno
- TCP Reno is optimized for single packet loss in a
window, multiple pkt loss degrades the
performance - TCP Newreno enhances the performance in the event
of multiple pkt drop without having SACK - Introduces new variable Partial ACK
- When fast retransmit is invoked record the
highest sequence number transmitted as Recover - When all of the data up to and including Recover
is acknowledged TCP exits fast recovery - Avoids invocation of Fast retransmit in quick
succession
12 Sender
Receiver
0 .. 61
14
7 .. 14
7 .. 13
cwnd 8
New round
14 dup Acks
0 .. 131
15 .. 28
15 .. 28
14
cwnd 15
New round
14, 29 .. 34
14
29 .. 34
cwnd 10
cwnd 21
Receiver generates Ack 29, sender exits Fast
recovery and adjust inflated cwnd and enters
Congestion avoidance
0 .. 281
cwnd 7
13TCP SACK
- SACK Data structure
- Recover and Partial ack as Newreno
- Receiver informs non-contiguous data
- Receiver can reflects max of three SACK blocks
- Sender maintains a queue of segments that have
been transmitted but not yet acknowledged,
Scoreboard - SACK is invoked in the event of three dup_acks
- Sender can judiciously send new packet or
retransmit packet to fill hole at receiver
14TCP SACK Data Structures
- struct tcpcb
-
- struct sackhole snd_holes
-
-
15SACK Example
Senders queue
Receivers queue
1
2
3
4
5
6
1
2
ACK 1 SACK 2
1
3
4
5
6
2
3
4
ACK 1 SACK 3,2
1
4
5
6
6
2
3
5
ACK 1 SACK 5,3,2
1
4
6
HOLES AT RCVR
16Evolution of ACK received by TCP-SAC Sender
3 dup_ack
Sequence Number
Time
17TCP Protocols Summary
Partial ACK Recv holes details Modified Recv
RENO NO NO NO
NEWRENO YES NO NO
SACK YES YES YES
18Outline
- Motivation
- TCP Mechanism
- Energy Model
- Experiment Methodology
- Experiment Results
- Conclusion
19Energy consumption of a node
Packet Reception Processing Packet
Processing Transmission
Idle Period
Current
PTX
PRX
Idle Current Consumption Pidle
Time
20Energy Model
- Total energy (E) can be represented as
- PTX is energy spent for transmission
- tTX is time spent in transmission
- tRX is negligible so can be eliminated
- Assume average connection throughput is ?
bytes/sec and the transmission speed is r
bytes/sec we can write
21Goals
- Awake node during ideal time consume unnecessary
power - Power off nodes during idle period this gives
lower bound of Total Energy consumption though
not realistic call it Idealized Energy - Alternatively during idle period node is asleep
and in the event of a pkt node is awoken up by
radio
22Goals contd.
- Investigate the Energy consumption of TCP
variants (Total Energy vis-à-vis Idealized
Energy) - TCP connection characteristic in wireless
revisited - Low bandwidth connection
- Higher BER
- Mobility
- These leads to longer Idle time
- Hence TCP sender can sleep!!
23Evolution of TCP Sequence
Idle connection
Sequence Number
Time
24Outline
- Motivation
- TCP Mechanism
- Energy Model
- Experiment Methodology
- Experiment Results
- Conclusion
25Test Bed
- Three hop wireless ad hoc network
- Each laptop equipped with Lucent 11 Mbps WaveLAN
card Running FreeBSD 4.3 - SACK is implemented based on RFC 2018
- Dummynet used for packet loss and delay
- Energy measurement at sender using HP 34401A
multimeter and VeePro software
26Energy Measurement
- Two multimeters simultaneously used for measuring
Radio and Total Node Power - Multimeter takes current measurement with a
granularity of 1 millisecond - Multimeter dumps data to different laptop
- Integration of current voltage over connection
time gives total Energy
27Workload, Factors, and Metrics
- TTCP used as a constant workload generator - 5 MB
- Error 1, 5 and 10
- RTT 15, 40, 70, 100 and 130 msec
- Reordering 1 and 5
- Burst Error 85 for 1 sec every 12 sec
- MTU 512 and 1500
28A typical run ..
Pkt
Router generates Loss and Delay (DUMMYNET)
Pkt
Start TTCP
Start LossDelay script
Pkt
Start TTCP server
Start multimeter
Stop
29Energy Consumed by Node
Current (mA)
Idealized Energy
Simultaneous measurement of Energy consumed by
the radio
Time (msec)
30Metrics
- Total Energy/bit ( E ) measured in Joules/bit.
This includes the energy consumed while the
sender is idle - Idealized Energy/bit ( EI ) measured in
Joules/bit. This measure excludes the idle time
energy and thus more closely approximates the
cost of the various protocols - Goodput in kbps
31Outline
- Motivation
- TCP Mechanism
- Energy Model
- Experiment Methodology
- Experiment Results
- Conclusion
32 Experiment Results Burst Loss
33Burst loss
Energy
Goodput
NEWRENO-Total
Sack-Total
SACK
Goodput (kbps)
Energy (Joules e-6/Bit)
Reno
SACK-Ideal
NEWRENO-Ideal
Average RTT (msec)
Average RTT (msec)
34Random uniform Loss
351 Packet Loss
Total Energy
Idealized Energy
RENO512
SACK512
SACK512
Total Energy (Joules e-6/Bit)
Idealized Energy (Joules e-6/Bit)
SACK1500
RENO512
SACK1500
NEWRENO1500
Average RTT(msec)
Average RTT(msec)
36 5 Packet Loss
Total Energy
Goodput
RENO1500
NEWRENO1500
SACK512
SACK1500
RENO512 NEWRENO512
Total Energy (Joules e-6/Bit)
Goodput (Kbps)
SACK1500
SACK512
RENO512 NEWRENO512
RENO1500
Average RTT(msec)
Average RTT(msec)
37TCP Reno, Newreno and SACK (Thruput)
Average Thruput 5 Packet Loss, MTU 1500
SACK
Newreno
Thruput (bps)
Reno
Time
38 10 Packet Loss
Total Energy
Idealized Energy
SACK1500
SACK1500
Total Energy (Joules e-6/Bit)
Idealized Energy (Joules e-6/Bit)
NEWRENO1500
SACK512
NEWRENO512
Average RTT(msec)
Average RTT(msec)
39Loss Experiment Summary - (mtu 512)
1 Loss 5 Loss 10 Loss
Goodput SACK SACK SACK
Total Energy SACK SACK SACK
Idealized Energy Reno Newreno Newreno
40 Packet Reordering
41 Goodput Packet Reordering
1 Packet reorder 5
Packet reorder
SACK
SACK512 1500
1500
512
NEWRENO
NEWRENO
1500
Goodput (Kbps)
Goodput (Kbps)
1500
512
512
Average RTT(msec)
Average RTT(msec)
42 1 Packet Reordering
Total Energy
Idealized Energy
RENO512
NEWRENO512
NEWRENO512
RENO512
Idealized Energy (Joules e-6/Bit)
Total Energy (Joules e-6/Bit)
SACK512
SACK512
Average RTT(msec)
Average RTT(msec)
43 5 Packet Reordering
Total Energy
Idealized Energy
RENO
512
1500
NEWRENO
NEWRENO512
1500
512
Total Energy (Joules e-6/Bit)
Idealized Energy (Joules e-6/Bit)
NEWRENO1500
SACK512 1500
SACK512 1500
Average RTT(msec)
Average RTT(msec)
44Discussion
- SACK outperforms in most of the cases (Burst
Loss, Random uniform Loss and Packet Reordering) - Does SACK help YES/NO
- Higher Goodput at the cost of extra processing
- SACK has higher Idealized Energy
45Discussion contd.
Total Energy vs Goodput
- Energy a 1 / Goodput
- Idealized Energy is a
- factor of Protocol
E a 1/?
Energy (Joules e-6/Bit)
Goodput (Kbps)
46Discussion contd.
Timeouts (Loss case)
- At higher loss all protocols perform
- alike
- SACK512 leads to lesser timeouts
- SACK - No timeout in reordering
- case, Reno gt Newreno
SACK1500
Average Number of Timeouts
RENO512
NEWRENO512
SACK512
Loss
47Related Work
- TCP Energy based on simulation
- We measured energy using a wireless test-bed and
a real TCP/IP stack (FreeBSD) - Metrics considered were Total Energy, ratio of
successful transmission to total number of
transmission - Total as well as Idealized energy (Processing
energy)
48Conclusion
- Awake Energy a 1 / Goodput
- Idealized Energy is a factor of Complexity of the
Protocol - Channel conditions are important in selection of
particular Protocol - SACK is not suitable for nodes with low Idle
Energy - SACK has lower Total Energy and Idealized Energy
under Packet reordering
49Questions??
Thank You