Is Lambda Switching Likely for Applications

1
Is Lambda Switching Likely for Applications?

• Tom Lehman
• USC/Information Sciences Institute
• December 2001

2
Context/Time Frames

• Far Future (10 yrs)
• New end system architectures which can
  source/sink 80Gb/s
• Large availability of fiber which can support
  gt1000s lambdas per fiber
• Near Future (5-10 yrs)
• Improvements in existing end system architecture
• Must live with/use large amounts of
  non-dispersion compensated fiber currently in
  place
• Following slides will focus on this time frame

3
Network Component Considerations

• Before we discuss network architectures,
  protocols, and features, lets evaluate what we
  can expect from the components in this time frame
• End Systems
• Routers Switches
• Optical capacity

4
End System/Router/Switch Performance (order of
magnitude)

• Currently
• 2 GHz processors, 128 bit gt 250 Gbit/s
  processing power
• 250 MHz SRAM, 128 bit gt 32 Gbit/s memory
  bandwidth
• 133 MHz PCIx, 64 bit gt 8 Gbit/s I/0 bandwidth
• Router/Switch Interfaces gt 10 Gbit/s
• Near Future?
• 10 Ghz processors, 128 bit gt Terabit/s
  processing power
• 1 GHz SRAM, 128 bit gt 128 Gbit/s memory
  bandwidth
• QDR 133 MHz PCIx, 64 bit gt 32 Gbits/s
• Infiniban12x gt 24 Gbit/s (per full duplex
  channel)
• Router/Switch Interfaces gt40 Gbit/s

5
Optical Capacity (order of magnitude)

• Currently
• C Band (1529-1562nm) _at_50-100Ghz spacing gt 64
  lambdas/fiber
• 10 Gbit/s/lambda gt 640 Gbits/s/fiber
• Near Future
• 80 Lambdas/fiber
• 40 Gbit/s lambda gt 3.2 Terabits/s/fiber
• 80 Gbit/s lambda gt 6.4 Terabits/s/fiber
• End Systems and Router/Switch interfaces will not
  keep up with the increase in optical speeds
• Moores law will not allow processor speeds to
  keep up
• End system architecture (complicated OS and
  protocol stack and slow I/O) will prevent
  matching speeds
• Router switch interfaces beyond 40 Gbit/s will be
  a challenge (electronic components do not yet
  exist for those speeds, so gap between optics and
  electronics will get larger)

6
Assumption on Component Capabilities

• End Systems
• Source/sink 24 Gbit/s
• End system modifications may include
• Direct host memory to NIC card dma features
• Complete offload of protocol stack to hardware
• Router/Switch Interfaces
• 40 Gbit/s
• Optical Capacity (per fiber)
• 80 lambda/fiber _at_40 Gbit/s gt 3.2
  Terabits/s/fiber

7
Is Lambda Switching Likely for Applications?

• No, in the context of end to end lambda switching
  between applications across the network on a
  widescale
• not enough lambdas for this to scale
• Special applications (e-Science?) may require end
  to end lambdas
• Yes, in the context of an interface or network
  connection which provides real time capacity
  provisioning based on application requests
• this could be part of overall QOS scheme
• (some) circuit/lambda switched in the core, IP on
  the edges
• Networks need to be able to support widescale
  deployment of fast end systems and also provide
  guaranteed QOS
• Lambda switching could be a key enabling
  technology

8
Network Issues

• Networks need to provide applications with on
  demand guaranteed QOS. This should include
• ability for applications to discover/query in
  real time end to end performance
• ability to be guaranteed a level of
  performance on and end to end basis
• IP Networks
• packet switched networks good for bursty traffic
  and applications which can live with best effort
  delivery.
• not so good for long flows which require
  guaranteed performance
• currently no well working mechanisms for
  guaranteed QOS in IP networks.
• DiffServ, RSVP, MPLS, IntServ
• Circuit switched networks provide QOS, but did
  not scale well.

9
GMPLS a solution?

• GMPLS may be heading in the right direction
• allows MPLS to control multiple devices such as
  LSRs, SONET ADMs, OADMs, OXCs
• Allows network to utilize benefits of packet
  switched and circuit switched networks.
• Bulk of traffic should still be packet switched
  on an end to end basis
• Circuit switched lambdas could be used for
  provisioning based on aggregation of application
  demands
• Special applications (e-Science?) may require end
  to end lambdas

10
GMPLS Architecture

• IP is the control plane for setting up layer 2
  circuits

IP Router
Lambda Switch
11
QOS in this Environment

• The packet switched QOS should be simpler
• just guarantees bandwidth?
• applications which use this can compensate for
  loss and jitter (including VOIP)
• Circuit switched QOS guarantees bandwidth, delay,
  jitter
• But the same issues that have yet to be solved
  regarding QOS still need resolution
• What is the economic model? Who pays?
• How is it enforced?
• The monitoring, enforcement, and accounting must
  be simple.
• How is this accomplished across domains?
• an ability to create layer 2 circuits across
  domains may help this