Comparison of TCP Performance Models

Transcription

1 Comparison of [July 23, 01] Hans-Peter Schwefel / Manfred Jobmann / Daniel Höllisch / Dan Heyman Siemens ICM / Technische Universität München / AT&T Labs Hans.Schwefel@icn.siemens.de

2 Motivation: Internet on Air (IoA) IoA: The Evolution Target of Mobile Networks and Internet Trends in Mobile Networks: y IP transport in the backbone; IP in RAN y Transport voice & data over IP y Terminate IP in the mobile host y Separation: Transport Control Mobile Networks Internet-on-Air Trends in the Internet: y Enable wireless access, support mobility y QoS beyond Best Effort y Security and AAA Internet TCP traffic dominant in the Internet Wireless TCP important in IoA Slide 2-7/23/2001

3 Challenge: Traffic/Performance Models for TCP [Even without wireless links] 3/94 943/ 3/ :8 3. 5,. 098 :8 3. 5, /-, ,; ,; , ,7,9 4397,11.24/0 805,7,9 4397,11.24/ / / ,11. 97, ,3.024/0 8706: 70/ , , / :,9 43 %!#034 %!&,88 &,88!,/ 0 09, $.422 ( %! ,309, 0 2,309, $ $ ( ( %!:789 24/ 1. 24/ 1.41$ $ ( Slide 3-7/23/2001

4 Network Scenario I: Multiplexed ON/OFF Traffic [without wireless link] connections M Scenario: N (LRD) ON/OFF sources, queueing/loss only at bottleneck router Models without flow-control known results in literature Here: Investigate average throughput per connection in TCP setting Two scenarios: (A) fast access, slow trunk, (B) vice versa Slide 4-7/23/2001

5 Network Scenario II: Parameter Settings Observed: Goodput per connection i: G i :=C i / D i Average Goodput & weighted average: where C i :=#packets in conn. i, D i :=duration of conn. i Slide 5-7/23/2001

6 Analytic Models I: TCP-UMass [Padhye, Firoiu, Towsley, Kurose] Approximation of throughput T in persistent TCP connection Derived from sender/receiver behavior Input Parameters b: # packets acknowledged by ACK [here: b=1] RTT: Average Round-Trip-Time (including queuing delay) W max : Maximum size of congestion window [here: W max =1024] T 0 : p: Average time-out interval Fraction of retransmitted packets Slide 6-7/23/2001

7 Analytic Models II: TCP-Engset [Heyman, Lakshman, Neidhardt] Modified Processor Sharing model on connection level: For j active users, each obtains service at rate ~λ p when jλ p ν ~ δν/j when jλ p >ν with attenuation factor δ 1 derived from TCP behavior, With R 0 = minimal Round-Trip Time = Σd i +1/ν+1/ λ p Note: # active connections can be represented by birth-death process Steady-state prob. independent of distribution of connection size ( no impact of Long-Range Dependence by PT conn. sizes) Slide 7-7/23/2001

8 Analytic Models III: TCP-NBurst [Schwefel] Source 1 Source 2 Source 3 Bottleneck Router + B ν [infinite queue] Cmp. TCP-Engset Packet-level extension of TCP-Engset model: Sharing of bandwidth: packet-rate λ p at source reduced to δν/j for j active sources when jλ p >ν conn. duration extended #packets in conn. unchanged Throttling only during congestion state when Q>=B Slide 8-7/23/2001

9 Comparison (I): Applicability of Models y TCP-UMass + Simple Formula, based on end-to-end TCP behavior -- Computation of throughput only -- Derived parameters required: RTT, T 0, p? assumes persistent connections y TCP-Engset + use of basic parameters only + dynamic connections (ON/OFF) -- no packet-level queue restricted performance parameters (e.g. loss rate)? Independence of connection size distribution (LRD properties) y TCP-NBurst + packet-level queue wide range of computable performance parameters -- computationally hard (matrix-algebraic methods), potential numerical problems Slide 9-7/23/2001

10 Comparison (II): Goodput per connection [Scenario (A)] Geometric Connection Size Power-Tailed Connection Size (LRD) TCP-UMass closest fit TCP-Engset and TCP-NBurst equivalent for large N (but underestimation) Slide 10-7/23/2001 Weighted Goodput >> Plain Average Excellent fit of TCP-UMass to weighted average TCP-NBurst not computable here

11 Comparison (III): Goodput per conn. [Scen. (A), N=20] Variation of Buffer-Size Variation of av. connection size Small buffers B<5: TCP-UMass good Large buffers: TCP-Engset and TCP- NBurst match plain average TCP-Engset best match for shape of curve Slide 11-7/23/2001 All models over-estimate throughput for short connections (initial slow-start) Good fit for n p >40

12 Comparison (IV): Scenario (B) Variation of # Sources, N Variation of Buffer-size, B Best fit of TCP-NBurst TCP-UMass over-estimates largely Goodput insensitive too large buffers TCP-UMass model useless Slide 12-7/23/2001

13 Summary & Outlook (Wireless scenarios) 425, ,3, 9. 24/0 8 %! %!&,88 &,88 %! %! ,3/,3/ %!:789 : :, :,9 4341%!#034 41%!#034 y %!&,8824/0 706: 708,// 9 43, , ,3.0 5,7,209078,8 35:9 y 24/0 81, /0 81, , , ,79 89, ,7/0/ 34970,7/0/ y!%/ 897 -:90/ / 897 -:90/ ,:80.,:80,5,5 0 90/,;07, 0 5, 3,;07, 0 5, 3,;07, /5:9 44/5: ,70706: 70/ :9 44 Г 3. :/ y /0, 8 / y /0 5, / / , ,3, 9.24/ Г 3;089,90 25,.9 41%!03, , $ $ $ $! Slide 13-7/23/2001