Localized Asynchronous Packet Scheduling for Buffered Crossbar Switches

1
Localized Asynchronous Packet Scheduling
forBuffered Crossbar Switches

• Deng Pan and Yuanyuan Yang
• State University of New York Stony Brook

2
Outline

• Introduction
• Related work
• Localized asynchronous packet scheduling
• Simulation results
• Conclusions

3
Introduction

• Crossbar switches have long been the preferred
  structures for high speed switches and routers
• Provide non-blocking capability.
• Overcome the bandwidth limitation of bus-based
  switches.
• Packet forwarding is simple.

4
Introduction

• For a crossbar switch, packets may be buffered at
  either
• Output ports
• Input ports
• Crosspoints

5
Introduction

• Output queued (OQ) switches only have buffer
  space at the output side.
• Achieve 100 throughput.
• Require speedup of N for an NxN switch.
• Input queued (IQ) switches only have buffer space
  at the input side.
• Require no speedup.
• Have to work with high time complexity algorithms
  in order to achieve 100 throughput.

6
Introduction

• Combined input-output queued (CIOQ) switches make
  a trade-off between the crossbar speedup and the
  complexity of the scheduling algorithms.
• Have small fixed speedup of two.
• Achieve 100 throughput with any iterative
  maximal matching algorithms.
• Emulate OQ switches.

7
Introduction

• Buffered crossbar switches are a special type of
  CIOQ switches.
• Each crosspoint of the crossbar has a small
  buffer.
• Crosspoint buffers eliminate the input and output
  contention.
• Buffered crossbar switches can directly schedule
  and switch variable length packets.

8
Introduction

• Previous scheduling algorithms for crossbar
  switches mainly focused on fixed length packet
  scheduling or cell scheduling.
• At input ports, new packets are segmented into
  fixed length cells.
• The cells are used as the scheduling units and
  transmitted across the switching fabric.
• At output ports, the cells are reassembled into
  original packets.

9
Introduction

• Variable length packet scheduling, or packet
  scheduling, improves the switch efficiency by
  avoiding the segmentation-and-reassemble (SAR)
  process.
• Higher throughput.
• Shorter packet latency.
• Lower hardware cost.

10
Introduction

• Turner Infocom06 proposed two packet
  scheduling algorithms for buffered crossbar
  switches.
• They can provide work-conserving guarantees, or
  emulate scheduling algorithms for OQ switches.
• They schedule packets by imposing an order on
  buffered packets.
• Each crosspoint needs 2L or more buffer space,
  where L is the maximum packet length.

11
Introduction

• We consider the other side of the problem, low
  time complexity and easy to implement packet
  scheduling algorithms.
• We present the Localized Asynchronous Packet
  Scheduling (LAPS) algorithm and analyze its
  performance.
• Local info based
• No comparison
• Crosspoint buffer size of L

12
Outline

• Introduction
• Related work
• Localized asynchronous packet scheduling
• Simulation results
• Conclusions

13
Related work

• Scheduling algorithms in the literature for
  buffered crossbar switches are generally designed
  with two possible objectives
• To achieve high throughput.
• To emulate scheduling algorithms for OQ switches.
• The latter is a stronger requirement, but the
  implementation of the former can be simpler.

14
Related work

• Cell scheduling algorithms for high throughput
• CIXB-1, CIXOB-k, MCBF, SCBF
• Cell scheduling algorithms to emulate scheduling
  algorithms for OQ switches
• GBVOQ_OCF, GBFG_SP, MCAF-LTF
• Packet scheduling schemes
• Packet VOQ, Packet LOOFA, DPFQ

15
Outline

• Introduction
• Related work
• Localized asynchronous packet scheduling
• Simulation results
• Conclusions

16
Localized asynchronous packet scheduling

• Structure of a buffered crossbar switch
• Ini input port
• Outj output port
• Bij crosspoint buffer
• Qij virtual queue
• The crossbar has
• speedup of two.

17
Localized asynchronous packet scheduling

• Based on the locations of the packets to be
  scheduled, there are three types of scheduling
  involved in a buffered crossbar switch.
• Input scheduling
• Crossbar scheduling
• Output scheduling

18
Localized asynchronous packet scheduling

• Output scheduling has been well studied, and
  various scheduling algorithms are proposed.
• Output scheduling usually does not affect the
  throughput performance as long as they are
  work-conserving.
• We use a simple FIFO algorithm for output
  scheduling, which is work-conserving.

19
Localized asynchronous packet scheduling

• For input scheduling,
• Select a backlogged virtual queue whose
  crosspoint buffer is empty, and send its head
  packet to the crosspoint buffer.
• When there are multiple eligible virtual queues,
  different arbitration rules can be used.
• Since the crossbar has speedup of two, the packet
  is sent to the crosspoint buffer with bandwidth
  2R.
• Crossbar scheduling is similar.

20
Localized asynchronous packet scheduling

• In order to reduce the packet latency,
  cut-through switching can be used on the
  crossbar.
• Similarly, cut-through switching can be used at
  output ports.

21
Localized asynchronous packet scheduling
22
Localized asynchronous packet scheduling

• In input scheduling, the scheduling candidates of
  an input port are only the virtual queues whose
  crosspoint buffers are empty.
• This restriction simplify the implementation by
  enabling one bit to represent the status of the
  crosspoint buffer.

23
Localized asynchronous packet scheduling

• With speedup of two, LAPS achieves 100
  throughput for any admissible traffic.
• Define Zij(t)Qij(t)Bij(t)
• If Bij is not empty at time t, ?kZkj(t) has a
  negative derivative.
• If Qij is not empty at time t, ?kQik(t) ?kZkj(t)
  has a negative or zero derivative.

24
Localized asynchronous packet scheduling

• Assume that the traffic arrives according to a
  Poisson process and the packet length follows an
  exponential distribution with mean M.
• Ini can be approximately modeled as an M/M/1
  system, and accordingly

25
Localized asynchronous packet scheduling

• Hardware implementation
• Only local info is necessary, and it is suitable
  for distributed implementation and highly
  scalable.
• Since no comparison is necessary, the arbiters
  can implemented by priority encoders, which can
  make fast decisions in hardware.
• Since each crosspoint buffer needs only L buffer
  space, it minimize the cost for the switch.

26
Outline

• Introduction
• Related work
• Localized asynchronous packet scheduling
• Simulation results
• Conclusions

27
Simulation results

• We have conducted simulations to verify the 100
  throughput of LAPS and to measure its delay and
  buffer requirement.
• We consider five different LAPS implementations
• Fixed priority (FP)
• Random (RD)
• Round-robin (RR)
• Oldest packet first (OPF)
• Longest queue first (LQF)

28
Simulation results

• In order to reflect the burst nature of real
  network traffic, we emulate the incoming traffic
  by a Markov modulated Poisson process.

29
Simulation results

• We considered both uniform traffic and
  non-uniform traffic.
• The packet length in the simulation is uniformly
  distributed between 50, 1500 bytes.
• We consider a 1616 switch, and each input port
  or output port has bandwidth of 1G bps.

30
Simulation results

• Throughput

31
Simulation results

• Average delay

32
Simulation results

• Maximum queue length

33
Outline

• Introduction
• Related work
• Localized asynchronous packet scheduling
• Simulation results
• Conclusions

34
Conclusions

• Due to the introduction of crosspoint buffers,
  buffered crossbar switches can directly schedule
  and transmit variable length packets.
• Packet scheduling algorithms avoid SAR and are
  more efficient than cell scheduling algorithms.
• Higher throughput
• Shorter latency
• Lower hardware cost

35
Conclusions

• We presented the Localized Asynchronous Packet
  Scheduling (LAPS) scheme.
• Local info based
• No comparison
• Crosspoint buffer of size L
• We theoretically proved that LAPS achieves 100
  throughput with speedup of two, and conducted
  simulations to verify the results.

36
Thank you!

• Questions?