Outline • Motivation • Simulation Study • Scheduled OFS • Experimental Results • Discussion MIT LIDS Optical Flow Switching Motivation • OFS reduces the amount of electronic processing by switching long sessions at the WDM layer – – Lower costs, reduced delays, increased switch capacity Provide specific QoS for advanced services MIT LIDS LIDS OFS Motivation (cont) Elect. Domain 1KB 1MB Optical Domain 10MB Flow Size 100MB Elect. Domain 1KB 1MB Optical Domain 10MB 100MB Flow Size -Internet displays a “heavy-tail” distribution of connections -More efficient optics => more transactions in optical domain (red line moves left) MIT Optical Flow Switching Study • • • • • Short-duration optical connections – – Access area Wide area Network architecture issues – – – Connection setup Route/wavelength assignment Goal: efficient use of network resources I.e. high throughput Previous work: “probabilistic” approaches – – Difficulty: high-arrival rate leads to high blocking probability Problem: lack of timely network state information Our proposed solution: Use of timing information in network – – Schedule connections Gather timely network state information This demonstration – – – – Demonstrate flow switching Demonstrate viability of timing and scheduling connections Investigate key sources of overhead High efficiency MIT LIDS Connection Setup Investigation • Key issue: – – – • Previous work – – – • How to learn optical resource availability? Distribution problem “Wavelength continuity” problem makes it worse Addresses issues one at a time Assumes perfect network state information Will these results be useful for ONRAMP, WAN implementation? This work – – Assesses effects of distributed network state information Models some current proposals MP-lambda-S ASON MIT LIDS Methodology • Design distributed approaches – – • Baseline flow switching architecture – – – • Requested flows from user to user Durations on order of seconds All-optical Simulate approaches on WAN topology – – • Combined routing, wavelength assignment Connection setup End-to-end latency (“time of flight” only) Approaches: Ideal, Tell-and-Go, Reverse Reservation Assess performance versus idealized approach – Blocking probability MIT LIDS LIDS Ideal Approach Illustration Assume: Flow Requested from A->B l-Changers l-Changers A C l-Changers B A C Optical Flow D Bidirectional Multi-fiber Link B “Tell” cntl packet D l-Changers Network Infrastructure MIT LLR Routing, Connection Setup LIDS Tell-and-Go Approach Illustration Assume: Flow Requested from A->B Available l: 2,3 A Available l: 2,3,4 C Available l: 1,2 Link-state Updates B Available l: 1,2,3 D Link-State Protocol MIT A C Optical Flow B “Tell” Packet Single wavelength D Connection Setup LIDS Reverse Reservation Approach Illustration Assume: Flow Requested from A->B A C B A C B Information Packets Reservation Packet Route Chosen by B D Route Discovery MIT D Route, Wavelength Reservation Simulation Description • Results shown as Blocking Probability vs. Traffic Intensity – Uniform, Poisson flow traffic per node • Fixed WAN topology • Parameters: – – – – – – – F = Number of fibers/link L = Number of channels/link K = Number of routes considered for routing decisions U = Update interval (seconds) = Average service rate for flows (flows/second) l = Average arrival rate of flows (flows/second) = Traffic intensity. Equal to l/ not utilization factor MIT LIDS Simulation Topology MIT LIDS Latency-free Control Network Results (1sec flows) Title: Creator: gnuplot Previe w : This EPS picture w as not saved w ith a preview inclu ded in it. Comment: This EPS picture w ill prin t to a PostScrip t prin ter, but not to other types of prin ters. RR: F=1, L=16, K=10 MIT TG: F=1, L=16, K=10 LIDS Control Network With Latency Results (1sec flows) Title: Creat or: gnuplot Prev iew : This EPS pict ure w as not s av ed w ith a preview inc luded in it. Comment: This EPS pict ure w ill print to a Pos tSc ript printer, but not to other ty pes of printers. TG, RR: U=0.1, F=1, L=16, K=10 MIT LIDS Interesting Phenomenon • Why is TG performance better than RR? – 1 sec flows and large rho => small inter-arrival times Smaller than round trip time – – Thus, with high probability, successive flows will see same state (at least locally) Increases chance of collision Effect of distribution (latency) • Why is Rand better than FF? – – This is exactly opposite of analytical papers’ claim Combination of reasons Nodes have imperfect information FF makes them compete for same wavelengths (false advertisement) – Not seen in analysis because distribution was ignored MIT LIDS Scheduled OFS in ONRAMP • Inaccurate information hurts performance – – • Our proposal: Use of timing information to schedule flows – – – – • Deliver network information on time to make decisions Exchange flow-based information Maximize utilization of core network Possible small delay for user Issues – – – – • In this case: Simple speed of light Biggest problem: Core network resources wasted Can timing be implemented cheaply, scaled? Can schedules be implemented? Must make use of current/future optical devices Low cost ONRAMP OFS – – Demonstration of scheduled OFS in access-area network One example of an implementation MIT LIDS LIDS Scheduling in ONRAMP Intermediate Node Router OXC Sched OXC Access Node #2 Access Node #1 Control Control OXC OXC IP IP Router Router X-a R-a OXC Sched IP Fixed l Xponder FLOW Tunable l Xponder GE GE • OXC Sched FLOW IP Tunable l Xponder Fixed l Xponder GE Receiver (R ) )Xmitter (X) MIT GE ONRAMP Connection Setup • • Uses timeslotting and schedules for lightpaths X => li busy on output of node i at corresponding slot OXC Schedule Slot 1 l1 X l2 MIT X X l3 l4 Slot 2 Slot 3 X ….. LIDS LIDS Algorithm Timeline Overhead - Dependent on timing uncertainty Slot 1 Slot 2 Slot 3 TIME Scheduling OH Cannot go in next timeslot Scheduling OH Can go in next timeslot -Overheads includes all timing uncertainty -Efficiency of any scheduled algorithm related to timing uncertainty, and switching/electronic overheads -Rough efficiency = Flow duration / Flow duration + Overhead MIT Utilizing Link Capacity • Sending GigE over transparent optical channel – – • Clock rate 1.244 Ghz Rate 8/10 coding results in raw bit rate of 995.2 Mb/s Payload capacity for UDP – Send MTU-sized packets 9000 bytes Avoid fragmentation – Headers Ethernet (26 bytes) + IP (20 bytes) + UDP (8 bytes) = 54 bytes Result: 8946 bytes of payload/packet – Link payload limit 989.2288 Mb/s • Rate-limited UDP – – – Input: desired rate Timed sends of UDP packets achieve desired rates Demonstrates transparency of OFS channel MIT LIDS Experimental Setup • OFS implemented in lab • One second timeslots – • Routing/wavelength selection – – • All available wavelengths (currently 14) Both directions around ring Gigabit Ethernet link layer – • Timing overhead negligible Flows achieve theoretical maximum link rate ~989 Mb/s Rate limited UDP – – – – Unidirectional flows No packet loss (100s of flows) Variable rate Demonstrates transparent use of optical connection MIT LIDS OFS Performance Title: rate.eps Creator: gnuplot 3.7 patchlevel 1 Previe w: This EPS picture was not saved w ith a preview inclu ded in it. Comment: This EPS picture will prin t to a PostScrip t prin ter, but not to other types of prin ters. MIT Title: percent.eps Creator: gnuplot 3.7 patchlevel 1 Previe w: This EPS picture was not saved w ith a preview inclu ded in it. Comment: This EPS picture will prin t to a PostScrip t prin ter, but not to other types of prin ters. LIDS Current Performance Limitations Title: perc entcorr.eps Creator: gnuplot 3.7 patc hlev el 1 Prev iew : This EPS picture w as not s av ed w ith a preview inc luded in it. Comment: This EPS picture w ill print to a Pos tSc ript printer, but not to other ty pes of printers. MIT LIDS LIDS Current Performance Limitations(cont.) • Current overhead is 0.10 seconds – – Efficiency for one-second flows is therefore 90% Analysis of overhead reveals possible overhead of Gigabit Ethernet frame sync Still under investigation – – Switching overhead and timing uncertainty negligible I.e. scheduling viable, efficient Algorithm Overhead Timeline Flow begins………… 10ms 100ms 150ms time MIT Scheduling GBE Sync? Switching Command Receiver Laser