michal przybylski ( [email protected]) , artur binczewski ( [email protected]) ,
DESCRIPTION
Testing ATRIUM with demanding applications Application of the network traffic from the computation and storage grid environment for network performance evaluation from end-user perspective. Michal Przybylski ( [email protected]) , Artur Binczewski ( [email protected]) , - PowerPoint PPT PresentationTRANSCRIPT
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Testing ATRIUM with demanding Testing ATRIUM with demanding applicationsapplications
Application of the network traffic from the computation and storage grid Application of the network traffic from the computation and storage grid environment for network performance evaluation from end-user environment for network performance evaluation from end-user
perspectiveperspective
Michal Przybylski (Michal Przybylski ([email protected])[email protected]), ,
Artur Binczewski (Artur Binczewski ([email protected])[email protected]),,
Maciej Brzezniak ([email protected])Maciej Brzezniak ([email protected])
Miroslaw Czyrnek ([email protected])Miroslaw Czyrnek ([email protected])
Tomasz Rajtar ([email protected])Tomasz Rajtar ([email protected])
Maciej Stroinski Maciej Stroinski ([email protected])([email protected]),,
Szymon Trocha (Szymon Trocha ([email protected])[email protected]), ,
PSNC, POLANDPSNC, POLAND
Wim Barbaix (Wim Barbaix ([email protected])[email protected]), ,
ALCATEL BELL NVALCATEL BELL NV, , BELGIUMBELGIUM
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
ATRIUM (IST 1999-20675) ATRIUM (IST 1999-20675) „A testbed of terabit routers running MPLS over DWDM”„A testbed of terabit routers running MPLS over DWDM”
Key action: KA IV, Action Line: VII.1.2. RN2, Key action: KA IV, Action Line: VII.1.2. RN2, „testbeds for advanced networking and application experiments”.„testbeds for advanced networking and application experiments”.
June 2002June 2002 - - June 2003June 2003 (Extended to December 2003), b (Extended to December 2003), budget 8+ Mudget 8+ M€€
• AALCATELLCATEL
• FT R&DFT R&D
• FUNDPFUNDP
• PSNCPSNC
• TTELEFONICAELEFONICA
• ULGULG
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
New terabit router architectureNew terabit router architecture
– redundant hardwareredundant hardware
(4 SF, 2 RS, 2MS)(4 SF, 2 RS, 2MS)
– modular softwaremodular software (SOLARIS) (SOLARIS)
– high availability routinghigh availability routing
– scalable platformscalable platform
– scalable linksscalable links
– scalable performancescalable performance
– automated and fast provisioningautomated and fast provisioning
– cross layer traffic engineeringcross layer traffic engineering
– coordinated protection and recoverycoordinated protection and recovery
The Alcatel 7770 RCP is a scalable
carrier grade core platform,delivering differentiated
IP and MPLS services
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
European TestbedEuropean Testbed
RED IRIS
GEANTRENATER
PSNC
FT R&D
Alcatel
ULG
TID
STM-16
CCC 622Mbit/s
FUNDP
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
General approachGeneral approach
• All partners provide PC. PSNC provides necessary servers, All partners provide PC. PSNC provides necessary servers, HPCs, storage systemsHPCs, storage systems
• GEANT provides CCC connection for PSNC and TID; GEANT provides CCC connection for PSNC and TID; Necessary software is freely available or is provided by PSNCNecessary software is freely available or is provided by PSNC
• Different modes and different network parameters examinedDifferent modes and different network parameters examined
• TE and QOS modificationsTE and QOS modifications
• Real applications do not necessarily mean bandwidth Real applications do not necessarily mean bandwidth consumingconsuming
• Subjective/relative QoS metrics allowedSubjective/relative QoS metrics allowed
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Expected resultsExpected results
• Technology issues fully understoodTechnology issues fully understood– RouterRouter– NetworkNetwork
• PSNC and partners can understand GRID PSNC and partners can understand GRID applications behaviour with terabit router and large applications behaviour with terabit router and large networknetwork– Delay / jitter Delay / jitter – BufferingBuffering– ReorderingReordering
• Applications tuning to long fat pipeApplications tuning to long fat pipe
• GÉANT can understand how to provision, monitor and debug the CCC GÉANT can understand how to provision, monitor and debug the CCC serviceservice
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Demanding applications...Demanding applications...
GRID FTPGRID FTP GAMESSGAMESS VIDEOVIDEO
Max bw.Max bw. <250 Mbit/s<250 Mbit/s ?? <1.5Mbit/s<1.5Mbit/s
DelayDelay important for important for lots of small fileslots of small files
?? ??
LossLoss low enough to low enough to sustain TCP sustain TCP
sessionsession
?? ??
JitterJitter ?? ??
• Empty networkEmpty network
• Congested networkCongested network
... in different network scenarios... in different network scenarios
• QoS enabled networkQoS enabled network
• Faults and recoveryFaults and recovery
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
GRID FTPGRID FTP
• Globus Toolkit v.2.0Globus Toolkit v.2.0• FTP based protocol working in 4 modes:FTP based protocol working in 4 modes:
– Stream mode (raw data in the data connection + ctrl stream)Stream mode (raw data in the data connection + ctrl stream)
– Extended block mode (file divided into blocks)Extended block mode (file divided into blocks)
– Parallel data transfer (1*ctrl + n*data, increase efficiency in Parallel data transfer (1*ctrl + n*data, increase efficiency in
long latency networks)long latency networks)
– Stripped data transfer (n*ctrl + n*data + n*server, increase Stripped data transfer (n*ctrl + n*data + n*server, increase
efficiency in long latency networks)efficiency in long latency networks)
• manual control of TCP buffer sizemanual control of TCP buffer size
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
GRID FTP measurementsGRID FTP measurements
– Total experiment timeTotal experiment time
– Total „data transmission” timeTotal „data transmission” time
– Avg., max., min. transmission bandwidthAvg., max., min. transmission bandwidth
– Time necessary to perform control operationsTime necessary to perform control operations
– Application usabilityApplication usability
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
GAMESS - The General Atomic and GAMESS - The General Atomic and Molecular Electronic Structure SystemMolecular Electronic Structure System
• Wavefunctions computationsWavefunctions computations• Designed for parallel computers & clustersDesigned for parallel computers & clusters• Able to run in distributed environment (TCP)Able to run in distributed environment (TCP)• Massive communication (peer to peer)Massive communication (peer to peer)
– input datainput data– computation resultscomputation results– computation synchronisationcomputation synchronisation
Measurements:Measurements:• time necessary to finish computationstime necessary to finish computations• application usabilityapplication usability• validity of the resultsvalidity of the results
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Multimedia streamingMultimedia streaming
• Proprietary multimedia distribution systemProprietary multimedia distribution system• Based on the independent node conceptBased on the independent node concept• Windows Media ServicesWindows Media Services• Live streamingLive streaming• On-demand contentOn-demand content• Multicast transmissionMulticast transmission
MeasurementsMeasurements– Subjective video qualitySubjective video quality– Available/used bandwidthAvailable/used bandwidth– Packet lossPacket loss– Frame lossFrame loss– RetransmissionsRetransmissions
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Delay / jitter on empty network [ms]Delay / jitter on empty network [ms]
FT
ALCATEL
FUNDP
3.5/0.1
1.5/0.2
2.5/0.1
33/0.3
PSNC1
PSNC2
36/0.3
35/0.3
1/0.1
100M622M1G
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Empty network: UDP/TCP bw [Mbit/s]Empty network: UDP/TCP bw [Mbit/s]
FT
ALCATEL
FUNDP
90/60650/100 PSNC1
PSNC2
95/7
460/8
350/240
91/90
455/9
100M622M1G
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
The more TCP sessions, the better results The more TCP sessions, the better results (up to a certain limit)(up to a certain limit)
GRID FTP – parallel transmissionGRID FTP – parallel transmission
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
Increase of the number of TCP sessions works Increase of the number of TCP sessions works only for long latency linksonly for long latency links
GRID FTP – parallel transmission...(1)GRID FTP – parallel transmission...(1)
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
File size matters as well...File size matters as well...
GRID FTP – parallel transmission...(2)GRID FTP – parallel transmission...(2)
T N CT N C 2 0 0 3 2 0 0 3 Z a g r e bZ a g r e b, , M a y 1 9 – 2 4M a y 1 9 – 2 4
ConclusionsConclusions
• No routing/packet loss/delay/jitter problems on the network No routing/packet loss/delay/jitter problems on the network composed of 7770RCP router, dedicated optical paths and composed of 7770RCP router, dedicated optical paths and GEANT CCC connectionGEANT CCC connection
• For maximum performance GRID have to use parallel TCP For maximum performance GRID have to use parallel TCP transfers; Optimal application configuration depends on local transfers; Optimal application configuration depends on local conditions – file size, link delay, TCP windowconditions – file size, link delay, TCP window
• GEANT CCC connection keeps jitter low and does not increase GEANT CCC connection keeps jitter low and does not increase significantly the delay – a sufficient equivalent of dedicated significantly the delay – a sufficient equivalent of dedicated channel (for high cost!)channel (for high cost!)– Service monitoring issueService monitoring issue
• Testbeds must be hack-protected – 1 month delay due to Testbeds must be hack-protected – 1 month delay due to testbed hack – need to reinstall GRID software and firewall testbed hack – need to reinstall GRID software and firewall systemssystems