Effect of Background Traffic on Performance of High-speed TCPs .


67 views
Uploaded on:
Description
Effect of Foundation Movement on Execution of Rapid TCPs. Injong Rhee http://www.csc.ncsu.edu/personnel/rhee/North Carolina State College Teammates: Sangtae Ha, Lisong Xu, Long Le. Microsoft Workshop. Foundation. Try different things with linux 2.6.19 Iperf (1 TCP-SACK stream)
Transcripts
Slide 1

Effect of Background Traffic on Performance of High-speed TCPs Injong Rhee http://www.csc.ncsu.edu/workforce/rhee/North Carolina State University Collaborators: Sangtae Ha, Lisong Xu, Long Le Microsoft Workshop

Slide 2

Background Experiment with linux 2.6.19 Iperf (1 TCP-SACK stream) 1Gbit spine interface: NC (USA) – Korea – Japan (uncommon on account of research group in Japan) Slow window development of Reno-style TCP brings about under-use Korea NC 202ms Japan 48ms

Slide 3

High-Speed TCP Variants Many High-speed TCP variations have been proposed How would we be able to assess these conventions? Which criteria? CUBIC H TCP-Africa Compound TCP HSTCP BIC TCP-AReno FAST Scalable New Protocol TCP-Westwood

Slide 4

HTCP HSTCP STCP BIC Window Size CUBIC Time Window development designs HSTCP H-TCP Scalable BIC-TCP CUBIC NS2-Linux [?], 400Mbps, 160ms one-way delay,100% BDP support, No foundation activity

Slide 5

Performance Criteria and Design Tradeoffs There are numerous execution criteria Fairness Intra-convention reasonableness RTT-decency TCP-kind disposition Scalability (High connection usage) Stability Not all conventions fulfill every one of the objectives. Be that as it may, rather, make distinctive plan tradeoffs. For instance, abandon meeting time to acquire strength, or the other way around.

Slide 6

Performance Evaluation Methodology Internet explore Most reasonable tests, however Hard to duplicate the outcomes No thought on what occurred in the system Simulation or dummynet copying Easily reproducible and unquestionable Main issue: would they say they are sensible? how to reproduce the Internet conditions? Hypothetical examination Provide vital bits of knowledge on the conduct of conventions But helpful suspicions and less valuable for correlation (e.g., first request practices).

Slide 7

Testbed imitating - reproducing the Internet condition. Topology Can\'t display the many-sided quality of the whole system. In this manner, most assessments concentrate on one or a couple bounce conditions (or dumbbell). Workload To adjust, concentrate on infusing reasonable foundation activity into the bottleneck interface. As arriving streams more likely than not experienced many jumps, mirroring the movement design found in one center switch has some impact of imitating the topology. Not immaculate as it doesn\'t permit us to see the practices of conventions under numerous bottlenecks. Be that as it may, this can be overcome by utilization of a "stopping" part topology expecting bottleneck connections are just a couple.

Slide 8

Realistic foundation movement Hard to demonstrate its authenticity, however we can make at any rate the insights comparable. Measure the Internet activity in one Internet connection and concentrate its factual examples, for example, stream sizes, entry rates, transmission rates, and so forth. Exceedingly definite diversion of Internet activity (in light of these factual examples) are conceivable. Apparatuses: HARPOON, Tmix, and so forth. A no-nonsense way: simply imitate the examples for the most part saw in the Internet. Landings - exponential, overwhelming tail Flow sizes - a shifted type of substantial tail (distinctive body and tail) RTT varieties - log-typical

Slide 9

Our work We concentrate the effect of foundation movement designs on the execution of conventions. Imperative to comprehend their practices in the Internet-like situations. This will shed lights on various tradeoffs that distinctive conventions take.

Slide 10

Testbed (Dummynet) Setup Totally 18 servers for producing foundation movement and getting and sending convention streams. Foundation activity is pushed into forward/in reverse bearings Long-lived streams: Iperf , brief streams: Surge (web movement generator) The RTT of each stream is arbitrarily picked in view of info appropriation. Exploratory parameters: RTT (40ms to 320ms), cushion sizes (1MB to 8MB).

Slide 11

Five unique sorts of foundation activity Type I: Surge (LN Body 93%, Pareto tail 7%) Exponential entry (0.2) Type II: Surge (LN Body 70%, Pareto tail 30%) Minimum document measure for tail - 1MB Exponential landing (0.6) Type III: Type I (90%), P2P movement (10%) P2P movement - Pareto, Minimum 3MB Type IV: 100% log-typical body Type V: Type II + 12 extensive Iperf streams

Slide 12

Link usage and solidness No Background (Buffer 1MB) Type II (Buffer 1MB) Some conventions diminish use when the rate difference of foundation activity increments.

Slide 13

Link use, soundness and misfortune synchronization No Background Type II Utilization High-speed TCP streams Background activity High rate varieties of convention streams may bring about misfortune synchronization and low use.

Slide 14

Stability versus Interface use Protocol Stability (measured in CoV - Standard Deviation separated By mean) Link use

Slide 15

Link use and dependability under different activity sorts (HTCP) Link usage CoV

Slide 16

Fairness (measured in throughput proportion) TCP cordiality (RTT 42 ms; 2MB cushion) Intra-convention Fairness (RTT 82 ms) RTT-decency (stream 1: 42 ms; stream 2: 162 ms) Generally, H-TCP demonstrates the fabulous reasonableness paying little heed to movement sorts. All conventions enhance reasonableness with more fluctuation in bg movement, however the span of activity has the greatest effect (sort V).

Slide 17

TCP benevolence No foundation Type V Generally, all conventions enhance reasonableness with sort V foundation movement.

Slide 18

TCP-invitingness another look Type II movement with changing quantities of fast streams (320ms RTT). Measured the throughput of Type II activity. We don\'t discover much contrast in throughput.

Slide 19

Convergence speed Cubic H-TCP No foundation activity Type II

Slide 20

Conclusion Types of foundation movement uncover "the brute" in mask. E.g, Some conventions exchange joining speed for higher solidness. A few conventions exchange security for quicker union and reasonableness. Rate difference of foundation activity influences the solidness and furthermore connect usage. All conventions enhance decency and union speed with more foundation activity (estimate matters more than fluctuation).

Slide 21

Intra-convention decency No foundation (2 MB cushion) Type V (2 MB support)

Slide 22

Intra-convention reasonableness (FAST) Type I Traffic; 1 MB Buffer. Wrong estimation of least RTT causes distinctive streams to keep running at various rates

Slide 23

Link use v.s. cushion estimate As the cradle space builds, the dependability shows signs of improvement. 320ms RTT

Slide 24

Impact of support sizes Buffer size (1 – 8MB), four HS streams with the same RTT (320ms) As the cushion measure builds, the CoV of all conventions diminishes

Slide 25

Impact of clog Buffer size (2MB), two HS streams with the same RTT (40ms – 320ms), twelve seemingly perpetual TCP streams included Convex conventions have a huge varieties (raised requesting still exists)

Slide 26

NS2 Simulation Results (Loss Model)

Recommended
View more...