The Impact of Network Variabilities on TCP Clocking Schemes
-
Upload
multimedia-networking-and-systems-laboratory -
Category
Technology
-
view
526 -
download
5
description
Transcript of The Impact of Network Variabilities on TCP Clocking Schemes
![Page 1: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/1.jpg)
The Impact of Network Variabilities on The Impact of Network Variabilities on TCP Clocking SchemesTCP Clocking Schemes
Mar. 19, 2005
Kuan-Ta Chen, Polly Huang,
Chun-Ying Huang, Chin-Laung Lei
Department of Electrical EngineeringNational Taiwan University
Global Internet 2005
![Page 2: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/2.jpg)
2Global Internet 2005
OutlineOutline
Motivation
Why pacing could be more bursty?
The impact of network variabilities on the behavior of TCP clocking schemes
Conclusion
![Page 3: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/3.jpg)
3Global Internet 2005
TCP Clocking SchemesTCP Clocking Schemes
Self-clocking (a.k.a. ack-clocking)ACKs “self-clock” the data to the rate of the bottleneck link
Pacingresembles to a rate control mechanism but preserves the concept of window control
a common implementation: release a window of packets evenly within each round-trip time
In intuition, pacing will result in more smooth traffic, and smooth traffic will lead to better performance, however, …
![Page 4: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/4.jpg)
4Global Internet 2005
MotivationMotivation
Aggarwal, Savage, Anderson found pacing often results in lower throughput and higher latency.
We are motivated to evaluate ack-clocking and pacing schemes with more fundamental behavioral analysis, especially on the aspect of traffic burstiness.
![Page 5: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/5.jpg)
5Global Internet 2005
Our main resultsOur main results
Pacing traffic could be more bursty than ack-clocking traffic.
The comparative traffic burstiness of TCP clocking schemes are largely affected by network path properties
whether the round-trip times (RTT) are the same
the number of flows
Pacing is generally less bursty than ack-clocking with realistic settings, i.e., heterogeneous RTT flows.
![Page 6: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/6.jpg)
6Global Internet 2005
Why pacing could be more bursty?Why pacing could be more bursty?
Intuitively, pacing should be no more bursty than ack-clocking.
We shall illustrate why the phenomenon could happen by behavioral models.
![Page 7: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/7.jpg)
7Global Internet 2005
Behavioral models Behavioral models –– equal window sizeequal window size
Assumption: 3 flows, the same RTT, equal window size = 6
t: bottleneck service time for a packet
T/6 T/6 T/6
a packet train for each flow
packet trains are equally spaced in a RTT
![Page 8: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/8.jpg)
8Global Internet 2005
Behavioral models Behavioral models –– different window sizedifferent window size
Assumption: 3 flows, the same RTT, differentwindows size = 5, 3, 10, respectively.
T/5 T/3T/ 10
![Page 9: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/9.jpg)
9Global Internet 2005
The effect of window unThe effect of window un--synchronizationsynchronization
Generate packet arrival sequences by the behavioral models
T = 100 ms, t = 0.1 ms, 3 flows
compare two cases
synchronized windows: 30, 30, 30
un-synchronized windows: 20, 30, 40
Observe traffic burstiness based on the wavelet-based MultiResolution Analysis (MRA) for the synthesized traffic.
![Page 10: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/10.jpg)
10Global Internet 2005
The energy plotThe energy plot
Ack-clocking nearly remains its burstiness
Pacing become more bursty
The effect can be amplified by more flows (show later)
![Page 11: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/11.jpg)
11Global Internet 2005
Validation and SimulationsValidation and Simulations
Observation: window un-synchronization can raise burstiness of pacing traffic.
We conduct network simulations to:validate the observation
examine the impact of flow multiplexing
examine the impacts of other variabilities
![Page 12: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/12.jpg)
12Global Internet 2005
Simulation SetupSimulation Setup
the network simulator is ns-2
1--50 flows, RTT are fixed to 100 ms
network topology
x Mbps(bottleneck)s
1
N
4x Mbps
R
1
N
4x Mbps
![Page 13: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/13.jpg)
13Global Internet 2005
The Effect of Multiplexing The Effect of Multiplexing –– AckAck--clockingclocking
more bursty in small scales (still less bursty than Poisson)
much less bursty in large scales
![Page 14: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/14.jpg)
14Global Internet 2005
The Effect of Multiplexing The Effect of Multiplexing –– PacingPacing
burstiness raises in all sub-RTT time scales
due to the effect ofwindow un-synchronization.
![Page 15: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/15.jpg)
15Global Internet 2005
The Effect of Multiplexing The Effect of Multiplexing –– A ComparisonA Comparison
50 flows fl pacing is more bursty in most of sub-RTT time scales
the comparative burstiness of two schemes are very different with and without flow multiplexing
![Page 16: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/16.jpg)
16Global Internet 2005
Examine the effect of RTT heterogeneityExamine the effect of RTT heterogeneity
The simulation setup is almost the same except:
fixed to 50 flows
RTTs are drawn from an uniform distribution over (100 ms, 300 ms)
![Page 17: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/17.jpg)
17Global Internet 2005
The Effect of RTT HeterogeneityThe Effect of RTT Heterogeneity
Ack-clocking is much more bursty
mismatch of round trip times
ack-solicited pkts are no longer spaced by t
Pacing is unaffected
RTT/window is alreadyrandomized by unsynchronized windows
![Page 18: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/18.jpg)
18Global Internet 2005
More Network VariabilitiesMore Network Variabilities
Simulations with additional factors:multi-hop, two-way traffic, cross-traffic, and their combinations
ID Topology RTT Heter.
Two-Way Traffic
Cross Traffic
- -
-
TwoWay Dumbbell ✓ ✓ -
Cross Dumbbell ✓ - ✓
Real Parking-lot ✓ ✓ ✓
-
-✓
Dumbbell
VarRTT Dumbbell
Fixed
![Page 19: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/19.jpg)
19Global Internet 2005
Network Variabilities on AckNetwork Variabilities on Ack--clockingclocking
The heterogeneityin flows RTT is a deciding factor.
![Page 20: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/20.jpg)
20Global Internet 2005
Network Variabilities on PacingNetwork Variabilities on Pacing
None of variabilities significantly affect pacing’s behavior
As long as RTTs are heterogeneous:
Ack-cloking is no less bursty than Poisson
Pacing is no more bursty than Poisson
flPacing is less bursty
![Page 21: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/21.jpg)
21Global Internet 2005
ConclusionConclusion
Provided physical explanation for ‘why pacing could be more bursty than ack-clocking’
Comparative burstiness of the TCP clocking schemes are network condition dependent, especially RTT heterogeneity and flow multiplexing.
It’s critical to include sufficient variabilitiesin performance evaluation of TCP based protocols.
![Page 22: The Impact of Network Variabilities on TCP Clocking Schemes](https://reader034.fdocuments.in/reader034/viewer/2022051817/549308deac79593b2e8b4763/html5/thumbnails/22.jpg)
Thank You!Thank You!