Support for high performance UDP/TCP applications Xing Li 2007-08-25.

Support for high performance UDP/TCP applications

Xing Li2007-08-25

Background

• More and more demonstration and trials on high-performance applications– Application-specific stream or data– Consume large network bandwidth– Heavily rely on network support

• What can NOC do to support for it?– If there is a lambda path

• lucky!• RTT is still an issue

– Otherwise• Measurement and monitoring

– Network performance measurement tool– Application-specific measurement tools

• End to end performance guarantee– Inter-AS routing adjustment– Agreement on enabling application-related ACLs

Where are the bottlenecks?

AS2 AS2 AS3 AS3

AS1 AS1

access access

Bad cablingWrong speed

Wrong duplex mode

Bad cablingWrong speed

Wrong duplex mode

firewall firewallBandwidth bottleneck

Host parameter

s

Host parameter

s

Possible Paths

Case study

In KRIt is not via the CERNET – KR pathIt is not via the CERNET– APAN – KR pathIt is not via the CERNET – TEIN2 – KR pathIt is not via the CERNET – TEIN2 – JP – KR pathIt is via the CERNET – APAN – Abilene – KR path

And the return path may be totally different

The Ideas

• Use measurement tools to– Help users to do self-service– Separate the network problems from the end

system problems– Find the network segments which cause the

problem

SASM Framework

The operation modes of basic tools

Server discovery and registration

From Arlington to TEIN2 Beijing server

using WALN Not good

using RJ45 Good

TCP transfer speed

max__ _

_ _throughput

receive window size

round trip time

max_,

.. / secthroughput bytes

65 535

056117 027

max_,

.. / secthroughput bytes

24 567

05643886

Formula

SUN §WINDOWS

Theory

The TCP measurement Tool

The testing steps

• Default test– The default TCP transfer speed for the e2e peer

• Window size test– Check the client window

• Parallel session test– Check the server window– Physical speed limit

• Comparison with ping and UDP (dvping) tests– Check packet loss (ping and dvping)– Check RTT (ping and dvping)– Check throughput (dvping)

Comparison with UDP (dvping)

• Compare with UDP (dvping)– Check RTT, packet loss and the physical

speed limit

Summary of the TEIN2 NOC measurement servers

server clinet

ICMP

ICMP

U/UDP

U/UDPM/UDP

U/UDP

U/UDP

U/TCP

U/TCP

U/UDP

ping

ssmping

dvping

iperf –c

smjoin

iperf –s

smclock

dvmcast

ssmpingd

Socket stack

U – UnicastM - Multicast

202.179.252.102

202.179.244.105

202.179.242.102

202.112.35.200

CERNET

TEIN2

TEIN2 Measurement Server Locations

SASM Framework

A Document

http://noc.tein2.net:8036/support-for-hp/hp-mc-20070507.html

Login: tein2 passwd: tein2

eVLBI example

Summary

• Application specific tools– dvping (UDP)– iperf (TCP/UDP)

• Distributed servers alone the path – Scalable Application Specific Measurement (SASM)

• Open traffic monitoring– SNMP– Netflow

• Applications– TCP vs. UDP

• International collaboration