Authors: Károly FarkasPolly Huang

Balachander KrishnamurthyYin Zhang

Jitu Padhye

EidgenössischeTechnische HochschuleZürich

2Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Talk Outline

� Introduction� Design Space of

TCP/HTTP Parameters

� Simulation Experiments

� Live Experiments � Conclusions� Future Plans

3Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Introduction

Carrier A

Carrier B

ISP A

ISP B ISP C

ISP D

Internet

HTTP/1.0, HTTP/1.1Application Layer - HTTPApplication Layer - HTTP

Transport Layer - TCPTransport Layer - TCP Tahoe, Reno, NewReno, SACK, …

4Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Comprehensive studyComprehensive study

TCP/HTTP aspects ⇒ Web performanceTCP/HTTP aspects ⇒ Web performance

Motivations & Goals

� Provide a taxonomy� Metodology to study

the key facets� ns based simulations� Live measurement

NO!

How?TCP/HTTP aspects ⇒ Web performanceTCP/HTTP aspects ⇒ Web performance

Comprehensive studyComprehensive studyGeneric Performance Forecast

5Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Design Space

Taxonomy of Parameters

TCP and HTTP interactionHTTPTCP

which may influence Web performance:» user perceived latency» server load» network load

6Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

TCP Parameters� TCP variants

» Tahoe» Reno» NewReno» SACK

� MSS (Max. Segment Size)

� Delayed ACK (Acknowledgement)

� Timer granularity� Initial RTO

(Retransmission Timeout)

� ICW (Initial Congestion Window)

� Timestamp

7Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

HTTP Parameters� Connection management

» Parallel connections – multiple connections between client ⇔ server at the

same time

» Persistent connections– the connection stays open ⇒ some TCP overhead can

be avoided

» Pipelining– multiple requests without waiting for individual

responses ⇒ latency could be reduced

8Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

TCP and HTTP Interaction� Network condition

» diff. degree/distribution of packet drops ⇒ diff. recovery of TCP variants

� Persistent connections» effect depending on initial RTO, timer granularity

� Parallel connections» effect depending on initial CW, drop handling

� TCP implementations» memory/computation resource consumption

9Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Simulation Experiments� Simulation software

» ns-2 (Network Simulator) � Simulation scenario ⇒ 50 ms

1 Mbps10 ms

10 Mbps10 ms

10 Mbps

from SURGEmodel⇐

� Workload settings

10Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Simulation Experiments (2)

� Parameter space ⇒

� Other settings» duration of the simulations: 30 min» 3 loss levels (0.1%, 1%, 3%)» number of repetitions: 10 with diff. seed values» one parameter was changed at a time in every simulation

11Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Simulation Results� Page download time

» MSS ↑ ; delayed ACK: off ⇒» TCP variants ⇒» Persistency: off ⇒» Parallel connections ↑ (loss: 0.1%; 1%) ⇒

� First response time» MSS ↑ ⇒

» Persistency: off ⇒» Pipelining; Parallel connections ↑ ⇒

12Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Simulation Results (2)� Object download time

» MSS ↑ ⇒

» Pipelining: on ⇒» Number of parallel connections ⇒

� Server load» TCP parameters ⇒» Persistency: off » Pipelining: on ⇒» Parallel Connections ↓

13Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Live Experiments� Measurement software

» Networking stack– the Alpine user-level library

» HTTP Client software–httperf

» HTTP Server software–boa Web server

� Test environment» local environment with dummynet

14Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Measurement Results� Content for study

» canonical page: 54 images with emp. distribution» downloaded images: first 6, 12, 18, 54

� Results» Tahoe: data limited, small CW ⇒» SACK: large amount of data ⇒» SACK: high loss, small CW ⇒» MSS ↑ , delayed ACK: off ⇒» A good value for initial CW: 4

15Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

Conclusions� Impact of TCP parameters

» MSS ↑ , delayed ACK: off ⇒ �

» other parameters ⇒ �

� Impact of HTTP parameters» pipelining + HTTP/1.1 � ⇔ serial + HTTP/1.0 ∅∅∅∅

� Combined Impact of TCP & HTTP parameters» high loss, short flows, serial HTTP/1.1 ⇒ SACK �» long flows, serial HTTP/1.1 ⇒ Tahoe �» low loss, short flows ⇒ every TCP variant �

16Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002

�To investigate the combined effect when more than one parameter has been changed

�To extend the live experiments from local area to wide area

Future Plans

17Károly Farkas, farkas@tik.ee.ethz.ch 28th May, 2002