Construction of an SNU Supercomputer Grid and Study of Efficient - PowerPoint by alextt

VIEWS: 15 PAGES: 25

									Construction of an SNU Supercomputer Grid and Study of Efficient Grid Networking Schemes
Prof. Chongkwon Kim, Seoul National University [ckim@popeye.snu.ac.kr] 2004.12.2

2008-11-04

 SNU INC Lab

Outline
  

Objectives Plan & Results Scope
 

Construction of an SNU Supercomputer Grid Study of Efficient Grid Networking Schemes
 Remedy RTT-unfairness of TCP in Fast Long-Distance Networks  Available Bandwidth Measurement



Conclusions

2008-11-04

 SNU INC Lab

Objectives
 

Construction of an SNU Supercomputer Grid Propose Efficient Grid Networking Schemes


RTT-fair transport protocol in fast long-distance networks
 Remedy RTT-unfairness of TCP in Fast Long-Distance Networks



Fast convergence algorithm for available bandwidth measurement

2008-11-04

 SNU INC Lab

Plan & Results
소요개월

1분기 1 2 3 4

2분기 5 6 7

3분기 8 9 10

4분기 11 12

비고

주요내용 테스트베드 구축 Measurement 제안 TCP 성능 향상 시뮬레이션
논문 작성 보고서 작성

2008-11-04

 SNU INC Lab

SNU Supercomputer Grid


Installation

2008-11-04

 SNU INC Lab

SNU Supercomputer Grid


Test

• Globus Toolkit 설치 테스트 방법 : • grid-proxy-init 프록시 생성 여부 확인 • globusrun a r [호스트이름] gatekeeper에서 인증처리 여부 확인 • globusrun o r [호스트이름] '&(executable=/bin/date)' jobmanager 동작여부 확인 • MPICH-G2의 간단한 설치 테스트 방법 : • mpirun globusrsl cpi.rsl v mpi 테스트 • 위의 두가지 test가 성공하면 타 기관 test bed와 연동하는 것에 문제가 없다고 봄

2008-11-04

 SNU INC Lab

Study of Efficient Grid Networking Schemes

1.

2.

Remedy RTT-unfairness of TCP in Fast Long-Distance Networks Available Bandwidth Measurement

2008-11-04

 SNU INC Lab

Grid and Fast Long-Distance Networks




Grid networks such as DataGrid will essentially require large bandwidth. But Grid can be considered long distance networks

Transport protocols will be more important as the fast long-distance networks are more popularized
 SNU INC Lab

2008-11-04

Problem with TCP


1500 bytes packet, 100ms RTT, 10Gbp


Physically impossible
 Sending rate (packets/RTT)

w

1 .5 ` 0 .5 p

 require window size W = 83,333 packets  at most 1 drop every 5,000,000,000 packets

Scalability Problem  low utilization

2008-11-04

 SNU INC Lab

Protocols enhancing scalability


HSTCP


1e+6

Sending rate R (Packet/RTT)

Make the parameters  and  be functions of current window  More aggressive when congestion window is larger

Standard TCP HSTCP "Scalable_TCP"

1e+5

1e+4

1e+3



Scalability TCP



1e+2

Multiplicative increase and multiplicative decrease

1e+1

1e-6

1e-5

1e-4 Loss Event Rate

1e-3

1e-2

1e-1



Severe RTT unfairness problem



Competing flows with different RTTs consume vastly unfair bandwidth shares The slope of response function indicates RTT-unfairness  SNU INC Lab

2008-11-04

Protocol reducing RTT-unfairness
 

BIC TCP Binary search increase
Sending rate R (Packet/RTT)

1e+6

1e+5

Standard TCP HSTCP "Scalable_TCP" BIC

1e+4

1e+3

1e+2

1e+1

1e-6

1e-5

1e-4 Loss Event Rate

1e-3

1e-2

1e-1


 

RTT fairness


be the same as regular TCP at a low loss rate, but STCP at a high loss rate

Partially reduces RTT-unfairness at the expense of TCP-friendliness Scalability problem may occur at a low loss rate
 SNU INC Lab

2008-11-04

Goals


Main properties


Bandwidth Scalability
 Sustain high speeds without requiring unrealistically low loss rates  Recover from congestion without huge delays



TCP-friendliness
 Modified TCP never take bandwidth away from original TCP when a packet loss rate is high.

  

RTT-fairness
 Competing flows with different RTTs consume fair bandwidth shares

Fast fair share convergence Deployment without router involvement

2008-11-04

 SNU INC Lab

Our approach


Follows different response functions in terms of RTT
 

Has smaller throughput gap between different RTTs than other existing high-speed protocols Becomes RTT-fair and TCP-friendly
Standard TCP HSTCP BIC EIMD (40ms) EIMD (240ms)
1e+6 Standard TCP 40ms High Speed 40ms High Speed 240ms EIMD 40ms EIMD 240ms

1e+6

1e+5

1e+5

1e+4

Thoughput (Packets/seconds)

1e+4

1e+3

1e+3

1e+2

1e+2

1e+1

1e+1

1e-6

1e-5

1e-4 Loss Event Rate

1e-3

1e-2

1e-1

2008-11-04

 SNU INC Lab

1e-6

1e-5

1e-4 Loss Event Rate

1e-3

1e-2

1e-1

EIMD - exponential increase


SIMD modification


to perform properly under high-speed networks



Window control rule
 
11 / u , u  2s /(1  s) Increase : wt  R  wt   ( wt  w0 ) Decrease : wt   wt  wt



Alpha is calculated to follow proposed response function
(1 s ) / s

 ( A  RTT r )1 /(1s )    u  1  (u /(u  1))    

     (u  1)    

1/ u

w

1 1s  u s max

2008-11-04

 SNU INC Lab

Simulations


RTT fairness


Under 2.5Gbps bottleneck link

RTT ratio EIMD
BIC HSTCP


1 1.03
0.88 1.01

3 3.2
10.63 28.03

6 4.91
35.08 108.08

Bandwidth utilization of EIMD


Can utilize the available bandwidth effectively under high bandwidth networks Bottleneck(bps) 200M 2.5G 5G

Utilization(%)
2008-11-04

86.22

93.34

96.95

 SNU INC Lab

Simulations


TCP friendliness


The sum of long-lived TCP flows and background traffic indicates the degree of TCP friendliness
Long-lived TCP 100% Background High Speed f0 High Speed f1 Unsued

80%

60%

40%

20%

0% EIMD BIC HSTCP EIMD BIC HSTCP



EIMD is the most TCP friendly regardless of congestion window
 SNU INC Lab

2008-11-04

Available Bandwidth Measurement




Available bandwidth is the maximum rate in which sender can send data without congestion in present network situation. In overlay networks like grid network, available bandwidth information is very useful in data-transport.
 

Choosing a server which can achieve optimal performance Adjusting parameters of transport protocol to avoid network congestion

Internet

2008-11-04

 SNU INC Lab

Available Bandwidth Measurement


In data transport,




If sender doesn’t know available bandwidth, network may be under-utilized or congested. If sender knows available bandwidth, sender can know appropriate sending rate.

Congestion! Available Bandwidth = ??

Available Bandwidth = 20Mbps

Well - utilized

!

2008-11-04

 SNU INC Lab

Bandwidth Measurement Tools
Bandwidth Measurement

Capacity Measurement

Available Bandwidth Measurement

Bottleneck Link Placement

bprobe PBM ADR CapProbe

Single-hop model

Self-induced Congestion

BFind Pathchar Pathneck

Delphi IGI/PTR Spruce

TOPP PathChirp Pathload

2008-11-04

 SNU INC Lab

Purpose of this Research


Fast and Accurate Available Bandwidth Measurement



Fast Measurement is important to cope with dynamic network conditions.


If the measurement takes long time, measured values may be out-of-date.

2008-11-04

 SNU INC Lab

Key Idea


Based on single hop gap model, we send probing packet trains and analyze packet gaps.

 

Send several packet trains. Analyze receiving packet gaps to infer available bandwidth. p  Bi g i go  C
 SNU INC Lab

2008-11-04

Key Idea
 

We should find the point when gap difference becomes 0. During the measurement, adjust sending packet gaps to measure the available bandwidth quickly. (linear interpolation)
Gap Difference Gap Difference

0.00025 0.0002 0.00015 0.0001 0.00005 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

go  gi 

2 g 1  g diff diff

g g
1 i

2 i

gi  K

2008-11-04

 SNU INC Lab

Simulations


Measurement time is shorter than previous methods.
Num of Phase vs. Com peting Traffic (40m s) 30
Number of Phase

25 20 15 10 5 0 5 15 25 35 45 55 65 75 85 95 Percentage of Com peting Traffic (%) IGI Our m ethod

2008-11-04

 SNU INC Lab

Simulations


We improved measurement accuracy.
Available Bandwidth Measurement 25

Measured Available Bandwidth

20 15 10 5 0 5 15 25 35 45 55 65 75 85 95 Percentage of Competing Traffic (% )

IGI O ur method Available Bandwidth

2008-11-04

 SNU INC Lab

Project Results
 

Construct SNU Supercomputer Grid
Propose a TCP that eliminating RTT-unfairness in fast long-distance networks such as Grid Network



A novel Available Bandwidth Measurement Tool.



Faster and more accurate than previous methods. Available Bandwidth Measurement information can be used to optimize performance of data-transport in grid network.



We are now applying for a patent.
 SNU INC Lab

2008-11-04


								
To top