Dualresource TCPAQM for processingconstrained networks

1
Dual-resource TCP/AQM for processing-constrained
networks
INFOCOM 2006, Barcelona, Apr. 25, 2006

• Minsu Shin and Song Chong
• Department of EECS,
• KAIST, Korea

Injong Rhee Department of CS, NC State Univ., USA
2
Outline

• Motivation
• Processing-constrained network
• Dual-resource environment
• Objective
• Dual-resource fair allocation
• Dual-resource TCP/AQM (DRQ)
• DRQ objective
• DRQ implementation
• Simulation results
• Conclusion

3
Processing-constrained network

• Link bandwidth grows fast
• Advancement in optical network technology
• Over-provisioning as the solution to congestion
• The rise of in-network applications

Future network
Virus detection
Complexity increases
Duplicate data suppression
Data Transcoding
VPN IPSec
Web-switching
Firewall
Network address translation
Packet classification and filtering
IP forwarding
Traditional network
4
Dual-resource environment

• Both bandwidth and CPU can be a bottleneck
• Can existing congestion control (TCP) be applied?

Malicious user
Congestion
Switch
CPU
User
What is fair and efficient resource allocation?
5
Objective

• To propose a dual-resource fairness criteria
• Extend the proportional fairness to the
  dual-resource environment
• Provide fair and efficient resource usages
• To propose a dual-resource queue (DRQ)
• Active queue management (AQM) strategy
• Approximate dual-resource fairness for TCP
  sources
• Scalable Not maintaining per-flow states or
  queues
• Incrementally deployable No changes in TCP
  stacks

6
Single-resource fairness

• Only considering link bandwidth constraint
• Assume that network consists of L links

rate r1
Output Link
Tx
rate rS
Bl (Mbps)
Maximization problem
Low 99
7
Dual-resource fairness

• Considering both CPU and bandwidth constraint
• Network consists of L links and K CPUs

rate r1
rate rS
Maximization problem
8
Performance

• Single link case

rate r1
rate r4
w 1, 2, 4, 8
Dual-resource fair allocation
Single-resource fair allocation
9
Dual-resource fair rate

• Congestion price of resources
• CPU price ?, Link price p
• Increasing demand gt resource capacity
• Decreasing demand lt resource capacity
• Positive when the resource becomes a bottleneck
• Zero when not a bottleneck
• Fair rate
• Inversely proportional to the aggregate price

Weighted CPU price sum
Link price sum
10
Dual-resource TCP/AQM

• Extend dual-resource fairness to TCP network
• DRQ modifies RED algorithm

TCP Sender (w)
TCP receiver
Tx
CPU
Tx
B1 (Mbps)
B3 (Mbps)
C2 (cycles/s)
Packet drop with probability p1
p2
p3
Current TCP/AQM
a
TCP Sending rate
Our Goal
11
DRQ algorithm
TCP Sender (w)
TCP receiver
Tx
CPU
Tx
B1 (Mbps)
C2 (cycles/s)
B3 (Mbps)
packet drop with probability p1
p2
p3
Communication between resources is needed
12
DRQ algorithm
At link 1, mark packet with prob.
At CPU 2, mark packet with prob.
At link 3, mark packet with prob.
If already marked, then drop packet!
Red card!
Yellow card!
No explicit communication between resources!
13
DRQ-ECN implementation

• Three ECN cases
• ECN 00 Initial state
• ECN 10 Signaling-marked
• (No congestion notification)
• ECN 11 Congestion-marked
• (TCP source decreases its window size by half)
• DRQs ECN marking algorithm
• When a packet arrives
• if(ECN ? 11) set ECN to 11 with red-card
  probability
• if(ECN 00) set ECN to 10 with yellow-card
  probability
• if(ECN 10) set ECN to 11 with yellow-card
  probability

14
Performance evaluation

• Comparison partners
• RED-RED CPU and link queues use original RED
• Very cheap. Most of current network system
  architecture
• DRR-RED Scheduling CPU using per-flow queue
• Expensive approach. Similar architecture to
  current computing system

RED
RED
DRR
RED
Output Link
Tx
CPU
C (cycles/sec)
B (Mbps)
15
Single CPU and link case
40 TCP sources, which require different
processing (0.25, 0.50, 1.00, 2.00)
Varying CPU capacity
In DRQ, each follows fair rates.
Topology
Average throughput of each source
DRQ
16
Single CPU and link case

• RED-RED has much lower bandwidth utilization
• DRQ performance is comparable to DRR-RED

Comparison of bandwidth utilization
17
Impact of high processing flows
Insert a few high processing flows (w10.0)
DRQ and DRR-RED prevent their domination but
RED-RED doesnt
18
Multiple-link simulation(1)
Parking-lot topology, with various cross-traffic
DRQ follows theoretic fair rates very well.
Throughput of TCP/DRQ in multiple link simulations
19
Partial deployment

• Network edge
• Pushing complicated tasks to the edge of Internet
• DRQ can be initially deployed to the edge system
• Simulation topology

Source groups SG1 High processing SG2 low
processing Others Negligible processing
20
Partial deployment
Throughput of processing-constrained edge
Partial deployment is also beneficial to improve
efficiency
21
Conclusion

• Contribution of this paper
• Finding an efficient and fair allocation policy
  in the dual-resource environment
• Suggestion of the practical implementation
  guideline

22
References
Kelly 98 Rate control in communication
networks shadow prices, proportional fairness
and stability", J. of the Operational Research
Society, 1998
Mo 00 Fair end-to-end window-based congestion
control", IEEE/ACM TON 2000
Wolf 00 Commbench a telecommunications
benchmark for network processors", ISPASS 2000
Low 99 Optimization flow control I Basic
algorithm and convergence", IEEE/ACM TON 1999
Floyd 93 Random early detection gateways for
congestion avoidance", IEEE/ACM TON 1993
Low 03 A duality model of TCP and queue
management algorithms", IEEE/ACM TON 2003
Pappu 02 Scheduling processing resources in
programmable routers, IEEE INFOCOM 2002