Towards a Robust Protocol Stack for Diverse
Wireless Networks
Arun Venkataramani
(in collaboration with Ming Li, Devesh Agrawal, Deepak Ganesan,
Aruna Balasubramanian, Brian Levine, Xiaozheng Tie at UMass
Amherst)
UNIVERSITY OF MASSACHUSETTS, AMHERST • Department of Computer Science
Wireless network landscape today
Diverse wireless networks coexist
Adapt TCP/IP stack in different ways
WLAN
Mesh
Little mobility
Cellular
MANET
Some mobility
DTN
High mobility
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
2
Interconnecting diverse networks?
Vision: A simple robust protocol stack to
interconnect diverse wireless networks.
WLAN
Mesh
Little mobility
Cellular
DTN
MANET
Some mobility
High mobility
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
3
Why does TCP/IP not suffice?
E2E route disruptions cause unavailability
WLAN
Mesh
Little mobility
Cellular
DTN
MANET
Some mobility
High mobility
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
4
Why does TCP/IP not suffice?
Gap between what you buy vs. get
WLAN
Mesh
Little mobility
Cellular
DTN
MANET
Some mobility
High mobility
Advertised
capacity:
11Mbps
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
5
Principle: Design for high uncertainty
Observe what works at an extreme point
in the design space--always-partitioned
DTNs--for insights into a robust design.
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
6
Outline
Motivation
Block transport
Replication routing
Research challenges
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
7
1. E2E Transport
E2E rate control is error-prone
E2E retransmissions are wasteful
E2E Feedback
Loss Rate
RTT
Rate
Control
Source
Congestion? Link loss?
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
Dest
8
1. E2E Transport
E2E rate control is error-prone
E2E retransmissions are wasteful
E2E Retransmissions
P
Source
Redundant
Transmissions
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
Dest
9
2. Packet as Unit of Control
Channel access
Link layer ARQ
Listen Backoff
RTS/CTS
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
11
2. Packet as Unit of Control
Channel access
Link layer ARQ
Timeout Backoff
Timeout Backoff
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
12
3. Complex Cross-Layer Interaction
Link-layer ARQ/backoff hurts TCP rate
control
Rate
Control
Transport
Highly Variable RTT
Link
Link ARQ
Link ARQ
Link ARQ
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
13
Hop: Clean-Slate Transport Protocol
End-To-End
Hop-by-Hop
Packets
Blocks
Complexity
Minimize interaction
Block-switched networks: A new paradigm
for wireless networks, Li et al. [NSDI 2009]
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
14
Hop Design
Virtual
Retransmission
Backpressure
Multi-hop
Per-hop
ACK
Withholding
Micro-block
Prioritization
Reliable Block Transfer
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
15
Reliable Per-Hop Block Transfer
Mechanisms
Burst mode (TXOP)
Block ACK based ARQ
Benefits
Amortizes control overhead
B-SYN Request for B-ACK
B-ACK Packet bitmap
CSMA
TXOP
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
16
Hop Design
Virtual
Retransmission
Backpressure
Multi-hop
Per-hop
ACK
Withholding
Micro-block
Prioritization
Reliable Block Transfer
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
17
Virtual Retransmission (VTX)
Mechanism
Leverages in-network caching
Re-transmits blocks only when unavailable in cache
Benefits
Fewer transmissions
Low overhead
Simple
A
B-SYN
B-ACK
DATA
C
B
D
E
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
18
Virtual Retransmission (VTX)
Mechanism
Leverages in-network caching
Re-transmits blocks only when unavailable in cache
Benefits
Fewer transmissions
B-SYNVTX
Low overhead
Simple
A
VTX Timer
B-SYNVTX
B
B-SYN with VTX flag set
D
E
B-ACK
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
19
Hop Design
Virtual
Retransmission
Backpressure
Multi-hop
Per-hop
ACK
Withholding
Micro-block
Prioritization
Reliable Block Transfer
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
20
Backpressure
Mechanism
Limits outstanding blocks per-flow at forwarder
Hold
B-ACK
A
Source
B
C
B-SYN
B-SYN
B-SYN
D
Slow
E
Dest
Limit on outstanding blocks=2
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
21
Backpressure
Mechanism
Limits outstanding blocks per-flow at forwarder
Benefits
Improves network utilization
A
20Mbps
Source
B
10Mbps
C
D
20Mbps
E
Dest
20Mbps
20Mbps
F
1Mbps
G
Dest
Aggregate goodput without backpressure: 6Mbps
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
22
Backpressure
Mechanism
Limits #outstanding blocks per-flow at forwarder
Benefits
Improves network utilization
A
20Mbps
Source
B
10Mbps
C
D
20Mbps
Dest
20Mbps
20Mbps
F
E
1Mbps
Limit of Outstanding Blocks=1
G
Dest
Aggregate goodput with backpressure: 10Mbps
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
23
Hop Design
Virtual
Retransmission
Backpressure
Multi-hop
Per-hop
ACK
Withholding
Micro-block
Prioritization
Reliable Block Transfer
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
24
Ack Withholding
Mechanism:
Receiver withholds
all but one BACK
A
C
B
Benefit:
Low overhead
Less conservative
Simple
B-SYN
B-ACK
DATA
B-SYN
Withhold
B-ACK
B-ACK
DATA
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
25
Hop Design
Virtual
Retransmission
Backpressure
Multi-hop
Per-hop
ACK
Withholding
Micro-block
Prioritization
Reliable Block Transfer
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
26
Micro-block Prioritization
Mechanisms
Sender piggybacks small blocks to B-SYN
Receiver prioritizes small block’s B-ACK
Benefits
Low delay for small blocks
SSH
Sender
64B
Receiver
B-SYN
DATA
B-ACK
Sender
B-SYN
FTP
1MB
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
27
Testbed
20 nodes on the 2nd floor of UMass CS building
Apple Mac Mini
1.8GHz, 2GB RAM, Atheros 802.11 a/b/g card
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
28
Single-flow Single-hop Performance
1.2x
TCP
1.6x
Hop
28x
Hop achieves significant gains over TCP
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
29
Single-flow Multi-hop Performance
1.9x
TCP
2.3x
Hop
2.7x
Hop achieves significant gains over TCP
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
30
Graceful Degradation with Loss
Emulated link layer losses at the receiver
TCP
Hop
TCP breaks down at moderate loss rates
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
31
Scalability to High Load
30 concurrent flows
2x
Mean
goodput
20x
TCP
150x
Hop
Hop-by-hop
TCP
Hop is much fairer than TCP
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
32
Hop Performance Breakdown
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
33
Low Delay for Small Transfers
1 small transfer competing with 4 large
AP
Small transfer size (KB)
Hop lowers delay across all file sizes
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
35
Related Work
Fixing E2E rate-control
Separating loss/congestion [Snoop, WTCP, Westwood+, ATCP, TCP-ELFN]
Network-assisted rate control [ATP, NRED, IFRC, WCP]
Hop circumvents rate control
Backpressure
ATM, theoretical work [Tassiulas et al.]
Tree/chain sensor data aggregation [Fusion, Flush]
Reliable point-to-point transport [RAIN, CXCC, Horizon]
Hop reduces backpressure overhead using blocks
Batching
Common optimization at link [802.11e/802.11n, WiLDNet, Kim08,
CMAP], transport [Delayed-ACK, DTN2.5], and network [ExOR] layers
Hop leverages batching across layers
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
40
Outline
Motivation
Block transport
Replication routing
Research challenges
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
41
Replication routing
Replication reduces delay under topology
uncertainty
Y
P
P
P
Z
X
P
W
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
42
Routing as resource allocation problem
Problem: Which packets to replicate given limited
X
bandwidth to optimize a specified metric?
Y
RAPID Protocol (X,Y):
1. Control channel: Exchange metadata
2. Direct Delivery: Deliver packets destined to each other
Change in utility
3. Replication: Replicate in decreasing order of marginal utility
Packet size
4. Termination: Until all packets replicated or nodes out of range
DTN routing as a resource allocation problem,
Balasubramanian et al. [SIGCOMM 2007]
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
43
Utility computation example
Objective: Minimize average delay
Define U(i) = -(T + D)
T = time since created, D = expected remaining time to deliver
Simplistic assumptions
uniform exponential meeting with mean ¸
global view
Z
i
i
i
X
Y
D=¸
D = ¸/2
W
D =¸/3
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
44
Utility computation example (cont’d)
j
j
i
W
X
Y
Deadline of i < T
Z
Deadline of j = T1 > T
Metric: Min average delay
Metric: Maximize #packets
delivered within deadline
Replicate i
Replicate j
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
45
Deployment on DieselNet
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
46
Results based on DieselNet traces
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
47
Outline
Motivation
Block transport
Replication routing
Research challenges
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
48
C1: When to replicate?
relays
Replication useful under
X1
Topology uncertainty
Delay optimization
Light load scenarios
X2
src
Claim: Delay benefit of
replication is unbounded.
dst
Xk
Xi = r.v. for delay of path i
Forwarding delay = mini(E[Xi])
Replication delay = E[mini(Xi)]
WLAN
Mesh
Low uncertainty
Forwarding suffices
DTN
MANET
High uncertainty
Replication useful
UNIVERSITY OF MASSACHUSETTS AMHERST • Department of Computer Science
49