Characterizing the RealTime Behavior of Prioritized SwitchedEthernet

1
Characterizing the Real-Time Behavior of
Prioritized Switched-Ethernet
P. Pedreiras, R. Leite, L. Almeida   DET-IEETA,
University of Aveiro Aveiro, Portugal
Electronic and Telecommunications Dept.
University of Aveiro Aveiro, Portugal
2
Motivation

• Ethernet
• General-purpose protocol, standard in computer
  data communications
• Ethernet in real-time applications
• Strengths
• High-bandwidth, cheap, mature, easy integration
  with Intra/Internet
• Weaknesses
• Inefficient handling of short data packets, no
  support for adequate scheduling policies, and
  mostly a non-deterministic arbitration scheme
• Approaches
• Switched Ethernet has been regarded as the
  ultimate solution for RT communication on
  Ethernet.
• Is this true ???

3
Presentation Outline

• Motivation
• Anatomy of a switch
• Potential problems on switched Ethernet LANs
• Experiments
• Conclusions and future work

4
Anatomy of a switch

• Private collision domain for each port(no
  collisions)
• Different traffic classes (up to 8)
• Virtual LANs(traffic isolation)

• Supports multiple data paths several data
  packets may be simultaneously transmitted
• Switch memory holds packets when there is
  concurrent access to port(s)

5
Potential problems of Switches

• However
• Broadcasts (frequent in RT applications) penalize
  switch efficiency
• Switch components (switch fabric, CPU, memory)
  are stressed by traffic handling in any VLAN
• Number of classes not enough to support efficient
  sched. policies
• Moreover
• Some management techniques unsuited for RT
  applications.
• Flow-control acts on Ports not on message classes
• Limitations on broadcasts per port not broadcasts
  per message

6
Experiments

• Experiment objectives
• Switch latency
• Interference between traffic in different VLANs
• Interference between distinct traffic classes

• Exp. 1 Switch latency
• Network loaded with only 1 periodic message
• Experiment carried with an Hub and after with a
  Switch
• Latency increase around 10µs
• Value compatible with cut-through forwarding
  technique
• Negligible when compared with other sources of
  perturbation (e.g. blocking due to non-preemption)

7
Experiments

• Exp. 2 Interference between different VLANs
• Switch fabric and CPU load
• VLAN1 carrying only one RT message (monitored)
• VLAN2 overloaded with (NRT) messages of several
  sizes
• Adding VLAN2 traffic caused an increase of the
  average jitter of the monitored stream from
  1.09µs to 1.12 µs -gt Negligible
• Memory usage under strong overloads
• 2 low-priority RT messages added to VLAN1(avg.
  load 32)
• VLAN2 and VLAN3 overloaded with (NRT) mesgs.
  (sev. sizes)
• RT traffic in VLAN 1 experienced lost packets
  (0.45) memory overflow caused by the traffic in
  VLANs 2 and 3 propagates to VLAN 1, causing lost
  packets.

8
Experiments

• Exp. 3 Interference between different priority
  levels
• 1 HP message (monitored, 46 data bytes, 1ms
  period)
• 2/3 LP load messages (50-200 data bytes,
  0.2-0.6ms period)
• Case 1 60 of the 10 Mbps bandwidth load
• Observed period and jitter have expected values

9
Experiments

• Exp. 3 (continuation)
• Case 2 103.5 of the 10 Mbps bandwidth load

• Switch misbehavior
• bursty transmission of HP messages
• jitter higher than expected

10
Experiments

• Results summary
• Switch latency small
• VLAN mutual interference due to CPU/Switch fabric
  load very small
• Memory interference (both between VLANs and
  traffic classes in same VLAN) dramatic
  effect, even over high-priority traffic

11
Conclusions and future work

• Conclusions
• Just adding a switch is not enough to achieve RT
  behavior
• Switches only operate as specified with
  controlled loads
• Flow-control mechanisms implemented by Switches
  are unsuited for RT applications
• New/existing techniques/protocols should be
  evaluated
• Future work
• Improvements to the monitoring tool
• Run this battery of tests in several switches
• Analytical tools to model some of these effects