12.07.2015 Views

Commissioning and Using the 4 Gigabit Lightpath iNetTest ...

Commissioning and Using the 4 Gigabit Lightpath iNetTest ...

Commissioning and Using the 4 Gigabit Lightpath iNetTest ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

UDP Throughput London-100009000Stockholm8000Alcatel Metro Core ConnectMCC Flow control ONKernel 2.6.20-web100_pktd-plusMyricom 10G-PCIE-8A-R Fibrerx-usecs=25Coalescence ONMTU 9000 bytesMax throughput 4.05 Gbit/sNO Packet loss – flowcontrolSending host, 3 CPUs idleFor


UDP Jitter Stockholm-LondonAlcatel Metro Core ConnectMCC Flow control OFFKernel 2.6.20-web100_pktd-plusMyricom 10G-PCIE-8A-R Fibrerx-usecs=0Coalescence OFF8972 byte packets16.5 µs spacingPeak in packet spacing at16 µs as expectedO<strong>the</strong>r peaks at 10 <strong>and</strong> 51 µsN(t)N(t)700000600000500000400000300000200000100000010000001000001000010001001018972B w=16.5 sto-man_rx-usecs0_5Apr080 20 40 60 80 100 120 140Latency us8972B w=16.5 sto-man_rx-usecs0_5Apr080 20 40 60 80 100 120 140Latency usconnect • communicate •collaborate


NORDUnet install TSSTest Stockholm – Manchester4 <strong>Gigabit</strong>E<strong>the</strong>rnetPath over TSSNORDUnet4 Gig TDM/SDHLighpathover GÉANT2+4 Gig on E<strong>the</strong>rnetOptical TransmissionJANET & NNWconnect • communicate •collaborate


UDP Throughput Stockholm-Londonwith Alcatel TSS – 10GE – MCCKernel 2.6.20-web100_pktd-plusMyricom 10G-PCIE-8A-R Fibrerx-usecs=75Coalescence ONMTU 9000 bytesMax throughput 3.78 Gbit/s(Previously was 4.05 Gbit/s)Packet loss 10%at packet spacing for 4.096Gbit/swith a 8129 byte payload.Recv Wire rate Mbit/s% Packet loss70006000500040003000200010000100sto-gig6_rx75_TSS_21May080 5 10 15 20 25 30 35 40Spacing between frames us806040200sto-gig6_rx75_TSS_21May080 5 10 15 20 25 30 35 40Spacing between frames us1000 bytes1472 bytes2000 bytes3000 bytes4000 bytes5000 bytes6000 bytes7000 bytes8000 bytes8972 bytes1000 bytes1472 bytes2000 bytes3000 bytes4000 bytes5000 bytes6000 bytes7000 bytes8000 bytes8972 bytesManchester – StockholmNO packet loss at 4.096 Gbit/sconnect • communicate •collaborate


Packet loss Stockholm-Manchesterwith Alcatel TSS – 10GE – MCC12Send 1M packets from PC inStockholm8192 BytesAll cross <strong>the</strong> TSS cloud.All enter Alcatel MCCas 10 GE framesPacket loss was in MCC Card~10% at 4096 Mbit/s% Packet Loss10864200 1000 2000 3000 4000 5000Offered Rate Mbit/sClassic “bottleneck”performance.But WHY ??connect • communicate •collaborate


UDP Jitter Manchester-Stockholmwith Alcatel TSS MCC – 10GE – TSS – PCKernel 2.6.20-web100_pktd-plusMyricom 10G-PCIE-8A-R Fibrerx-usecs=0Coalescence OFF8192 Byte packetsAlcatel MCC is TDM 28 VC-4.Can only send packets to <strong>the</strong>TSSat 16 µs not at 10 Gbit/s.Peak packet spacing is~ 6 µs ie packets arriving at10 Gbit/sN(t)N(t)N(t)400000350000300000250000200000150000100000500000100000010000010000100010010700000600000500000400000300000200000100000016 µs8192B w=16 gig6-sto_rx0_19J un080 20 40 60 80 100 120 140Latency us8192B w=16 gig6-sto_rx0_19J un080 20 40 60 80 100 120 140Latency us8972B w=16.5 sto-man_rx-usecs0_5Apr08Sto-Man no TSS16 µs0 20 40 60 80 100 120 140Latency usSuggests packet bunching in<strong>the</strong> TSS networkconnect • communicate •collaborate


UDP 1-way delayManchester-Stockholm withTSS Alcatel MCC – 10GE – TSS – PCKernel 2.6.20-web100_pktd-plusMyricom 10G-PCIE-8A-R Fibrerx-usecs=0Coalescence OFF8192 Byte packetsStructure similar to interruptcoalescence but this is OFF1-waydelayusJitter as function of recvtimesee delays of ~ 130 µs<strong>the</strong>n packets at ~6 µsThese packets at ~ 10Gbit/sjiterus190001895018900188501880018750187001865018600140120100806040200W16 gig6-sto_rx0_16J un087558000 7558500 7559000 7559500 7560000receive time us.W16 gig6-sto_rx0_16J un087558000 7558500 7559000 7559500 7560000receive time us.This confirms packet bunchingin <strong>the</strong> TSS networkconnect • communicate •collaborate


Why do we get packet lossin Copenhagen?The Alcatel TSS delivers <strong>the</strong> packets over <strong>the</strong> 10GE link to <strong>the</strong>Alcatel MCC in bunches at line-speed (with suitable pauses toget 4Gbit/s).The SDH circuit over GÉANT from <strong>the</strong> Alcatel MCC runs at 28VC-4 (~4.2Gbit/s).The Alcatel MCC ingress buffering is only about 57 k Byte.These factors toge<strong>the</strong>r cause buffer overflow <strong>and</strong> <strong>the</strong> packetloss.connect • communicate •collaborate


NORDUnet & GÉANT change<strong>the</strong> TSS – MCC linkA work-around !VLBI uses SDH to GÉANTAnd buffering in <strong>the</strong> TSSNORDUnetTSS Cloudconnect • communicate •collaborate


The Current 4 <strong>Gigabit</strong> link28 VC-4 SDHLinksTSS to MCC4 Gig onE<strong>the</strong>rnetOver DWDMJANET & NNWconnect • communicate •collaborate


INETTESTconnect • communicate •collaborate


<strong>iNetTest</strong>: Simulink design10GE MAC core tenGbE0Inter-packet timeHistogram tenGbE0NetworkparametersCSRsTransmit & receiveevent time logInter-packet timeHistogram tenGbE110GE MAC core tenGbE1Jonathan Hargreaves JBOconnect • communicate •collaborate


<strong>iNetTest</strong>: Details<strong>iNetTest</strong> FPGA runs with a 200 MHz (5 ns) clocksTwo <strong>iNetTest</strong> units can be controlled over e<strong>the</strong>rnet IP from a PC<strong>iNetTest</strong> E<strong>the</strong>rnet IP address <strong>and</strong> MAC preset in code– last digit selected by jumperTen GE IP address, Port <strong>and</strong> Gateway can be user configuredAutomatic or manual ARPing. ARP tables for each port can beexamined<strong>iNetTest</strong> responds to <strong>and</strong> generates PINGUDP packets can be sent by firmware between two <strong>iNetTest</strong>sor between <strong>iNetTest</strong> <strong>and</strong> PCsUser selects <strong>the</strong> number of packets to send, <strong>the</strong>ir length,<strong>and</strong> <strong>the</strong> time between <strong>the</strong>miBOBs count received <strong>and</strong> transmitted packets, <strong>and</strong> store up to2048 time stamped events per portiBOBs generate histograms of <strong>the</strong> arrival time distribution forreceived packetsconnect • communicate •collaborate


<strong>iNetTest</strong> & eMerlinIN:Architecture for Operation &Controlconnect • communicate •collaborate


<strong>iNetTest</strong> & eMerlinIN: GUIconnect • communicate •collaborate


<strong>iNetTest</strong> & eMerlinIN: GUIconnect • communicate •collaborate


Testing Long Term StabilityUDP Throughput Onsala-JodrellUse <strong>iNetTest</strong> to send trials of 100M 8192 Byte packets.Spacing 16 µs, Trial takes ~27 min.Measure:Throughput 4.094 Gbit/sPacket loss ZerojitteriBoB tenGbE0 locked up5000Ons-J bo_<strong>iNetTest</strong>_13May09805000Ons-J bo_<strong>iNetTest</strong>_14May097WireRateMbit/s45004000350030002500200015001000500706050403020%Packetlos10WireRateMbit/s45004000350030002500200015001000500654321%Packetlos00000 200 400 600 800 1000 12000 200 400 600 800 1000 1200Time during <strong>the</strong> transfer minTime during <strong>the</strong> transfer minconnect • communicate •collaborate


Testing Long Term StabilityUDP Jitter Onsala-Jodrell<strong>iNetTest</strong>Plot <strong>the</strong> inter-packet arrival times for 3 trialsNo noticeable differencesSame distribution for tests over several weeksWireRateMbit/s5000450040003500300025002000150010005000Ons-J bo_<strong>iNetTest</strong>_14May09765432100 200 400 600 800 1000 1200Time during <strong>the</strong> transfer min%Packetlosconnect • communicate •collaborate


Testing Long Term StabilityUDP Jitter Onsala-Jodrell<strong>iNetTest</strong>NO change in JitterTail ~ 10 -3 downPeak at 16 µs, tail extends to ~70 µsconnect • communicate •collaborate


What did we learn?The network is very stableiBOB state machine can lock <strong>the</strong> E<strong>the</strong>rnet transmit & receive(?)Added a reset function to <strong>the</strong> FPGA.E<strong>the</strong>rnet switches do age <strong>the</strong>ir MAC tables (as expected) SO:<strong>the</strong>y loose <strong>the</strong> port of <strong>the</strong> destination iBOBflood packets at 4 Gbit/s to all ports in <strong>the</strong> VLAN<strong>iNetTest</strong> & on2jbo_rx send “data ACK” packets back to <strong>the</strong>source every n packet received.<strong>iNetTest</strong> <strong>and</strong> eMerlinIN operate as required.Signals from signal generator in Onsala have been sampled<strong>and</strong> transmitted to <strong>the</strong> Station board in <strong>the</strong> Jodrell correlator.connect • communicate •collaborate


State histograms from <strong>the</strong>e-Merlin Station Board88 MHz Signal, 1024 MHz clock, 2 bits, 1pol.Local eMerlinIN88 MHz Signal,1024 MHz clock,2 bits, 1 pol.eMerlinIN from Onsala-40 dBm -10 dBm -6 dBm-3 dBm 0 dBm 3 dBmconnect • communicate •collaborate


ANY QUESTIONS ?connect • communicate •collaborate

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!