ops: optical packet switches - 123seminarsonly.com€¦ · sep 8, 2006 h. harai (nict) 3 optical...
TRANSCRIPT
OPS: Optical Packet SwitchesOPS: Optical Packet Switches
Hiroaki Harai ([email protected])National Institute of
Information and Communications TechnologySep 8, 2006Optical Network Testbeds Workshop 3
Sep 8, 2006 H. Harai (NICT) 2
Why do we Need OPS?Why do we Need OPS?Internet Traffic in Japan: approx. 500 GbpsPeta-bps backbone future: doubled per year 500 Tbps in 10 yearsElectronic packet switch
Year 2004: Throughput 640Gbps (16x40 Gbps)Lightpath networks
Need fully meshed connections/ feasible?Need complex traffic engineeringImportant technology for bandwidth-assured applications
OPS networksProvide extremely high-throughput
Much larger bandwidth for switching (> 40 Gbps)O/E/O: 40Gbps 64 x 622 Mbps bus, SERDES
May need MPLS-like control (labels can be merged)Important to ubiquitous society
Sep 8, 2006 H. Harai (NICT) 3
Optical Packet SwitchingOptical Packet SwitchingData-path is all-optical (No O/E/O)
Switch, BufferIncrease data bandwidth
Label lookup (i.e. forwarding)Electronic parallel processing?Optical processing
Optical
Electrical
payloadpayload headerheader
SchedulingAvoid packet collision
Priority control
SchedulingAvoid packet collision
Priority control
ForwardingDetermine output port from the routing table
ForwardingDetermine output port from the routing table
SwitchingSwitch the packet
to the appropriate port
SwitchingSwitch the packet
to the appropriate portBuffering
Store the packetsin appropriate time
BufferingStore the packetsin appropriate time
Routing Make a routing table for forwarding procedure
Routing Make a routing table for forwarding procedure
payloadpayload headerheader
Nx1 BufferNx1 Buffer
Nx1 BufferNx1 Buffer
Nx1 BufferNx1 Buffer
Nx1 BufferNx1 Buffer
1xN LabelSwitch
1xN LabelSwitch
1xN LabelSwitch
1xN LabelSwitch
1xN LabelSwitch
1xN LabelSwitch
1xN LabelSwitch
1xN LabelSwitch
payloadpayload payloadpayload
Electronicserial
Optical
payloadpayload payloadpayload
Sep 8, 2006 H. Harai (NICT) 4
What should be Solved for OPS?What should be Solved for OPS?OPS
Increasing number of ports of optical switchElectronic: 16x16, 40Gbps 640GbpsOptical: 128x128, 160Gbps 20.48 Tbps
25 Waves 500 TbpsIncreasing speed of label lookup and buffer management
Wire-speed operationIncreasing number of labels looked-up
Several thousands (New L2 possibility)More (L3 switching)
Increasing buffer sizeAt least tens of fiber-delay-lines
Decreasing guard time between packetsSeveral nanoseconds
OPS Monitor/AnalyzerBit error / Optical packet error
Under developing in NICT
Sep 8, 2006 H. Harai (NICT) 5
OPS PrototypeOptical label lookupOptical bufferElectronic buffer management
Remaining Topics of This TalkRemaining Topics of This Talk
Sep 8, 2006 H. Harai (NICT) 6
NICTNICT’’ss 40Gbps40Gbps--based OPS Experimentbased OPS ExperimentN. Wada, H. Harai, F. Kubota, OFC 2003 (no. FS7).
Sep 8, 2006 H. Harai (NICT) 7
Optical Optical CCode based ode based UUltra ltra FFast ast LLabel abel PProcessingrocessing
Optical label has different modulation format with payload data
Optical label is physically distinguished from payload data
Optical hardware based label processing is available
Fully passive, ultra high-speed optical label processing
Payload data
Packet format
Header(label)
Replace to the optical code (label)
Sep 8, 2006 H. Harai (NICT) 8
Time Domain Optical Code ProcessingTime Domain Optical Code Processing----Measured Waveform at 8Measured Waveform at 8--chip, 200Gchips/schip, 200Gchips/s
Ref.) K. Kitayama, N. Wada, IEEE Photonic Tech. Lett., vol. 11, pp. 1689—1691, Dec. 1999.
10ps/div.
10ps/div.Auto-correlation
Cross-correlation
Sep 8, 2006 H. Harai (NICT) 9
Optical FiberOptical Fiber--DelayDelay--Line BufferLine Buffer
Different lengths of FDLs
Need at least tens of FDLsH. Furukawa, H. Harai, N. Wada, N. Takezawa, K. Nashimoto, T. Miyazaki, ”A 31-FDL BufferBased on Trees of 1x8 PLZT Optical Switches,” to be presented at ECOC 2006, no. Tu4.6.5, Sep 2006. Discard
0T2T3T4T
(B-1)T
BufferManager
Nx(B+1)switch
11
2233
Control signal
Optical packets
44
55
55
44
11
3322
2211
3344
55
Switch 1 out
Switch 2 out Time (2µs/div)
Inte
nsity
(a.u
.)In
tens
ity (a
.u.)
Buffer 1 out
Buffer 2 out
Packet collision!!
Output port
Avoidance of collision
Buffer 1Buffer 1
Buffer 2Buffer 2
LN-SW LN-SW LN-SW
LN-SW LN-SW LN-SW
Output Output portport
LN-SW
LN-SWLN-SW
LN-SW
LN-SW
LN-SW
Single Switch
Double Switches
Time (2µs/div)
1
1 1 1
SchedulerScheduler
Noise
Optical FDL Buffering at 160GbpsOptical FDL Buffering at 160Gbps
Source: N. Wada (NICT)
Sep 8, 2006 H. Harai (NICT) 11
HighHigh--Performance Buffer Management Performance Buffer Management for Optical Fiberfor Optical Fiber--DelayDelay--Line BufferLine Buffer
Establish practical-scale high-performance management for FDL buffer(1) Develop buffer management by parallel and pipeline processing◇ For number of ports, time complexity of each processor is O(1)◇ Parallel expansion of sequential (i.e. round-robin) scheduling◇ N-times higher throughput than sequential scheduling
(2) Confirm feasibility of support for 128x40Gbps packet switch by FPGA◇ 8 times higher performance than ASIC based router (16x40Gbps)◇ IP packet granularity (64byte or more; 10 Gpps), variable length
(3) Prototyping 8-port buffer management system
8-port buffer management hardware
P q
P41
P42
P43
P44
P45
P46
P47
P48
l1
l2
l3
l4
l5
l6
l7
l8
∆1
∆2
∆3
∆4
∆5
∆6
∆7
∆8
Parallel and pipeline buffer management (N =8)
P35
P36
P37
P38
P23
P24
P25
P26
P27
P28
P12
P13
P14
P15
P16
P17
P18
cf) H. Harai and M. Murata, IEEE/ACM Transactions on Networking, Feb. 2006.
Sep 8, 2006 H. Harai (NICT) 12
Performance ComparisonPerformance Comparison
16000/port
2 Gpps **
125 Mpps/port **
40 Gbps16
Electronic Router
10 Gpps4 MppsScheduling
31/port2/portBuffer
10 Gpps (at 40Gbps)
160 Gbps2 *
NICT’s Top Data(As of Sep, 2006)
2IN/OUT ports160 GbpsBit rate
800 Mpps/portLabel processing
NICT OPS Prototype
* Can scale with nanosecond optical switches** Estimated data: Assumption of wire rate processing of 40byte-packets
Sep 8, 2006 H. Harai (NICT) 13
ConclusionConclusionWe need high-throughput backbone network for ubiquitous societyOPS will provide extremely high-throughput
Switching bandwidth is not limitedBuffer size is increasingElectronic scheduling is fast
NICT has developed OPS but,Need more advanced devices (e.g., ns-switch) and systems
AcknowledgmentN. Wada, H. Furukawa of Photonic Network Group in NICT for valuable discussion, collaboration, and some slides in OPS
Thank you for your attentionThank you for your attention