Less than Best Effort

1
Less than Best Effort
QoS IP 2003, Milan (Italy), 24-02-03
Nicolas Simar, Network Engineer DANTE
2
Agenda

• LBE definition
• LBE description
• Application scenarios
• Queuing technique
• Experimental results on GÉANT
• without congestion
• with congestion
• Acknowledgements

3
Less than Best Effort
Definition

• Per Domain Behavior allowing a traffic class to
  make use of the unutilised bandwidth in the
  network.
• in case of competitions for resources, the LBE
  traffic will be discarded before any Best Effort
  or higher-class traffic.
• Use of DSCP 001000 - same as Internet2 Scavenger
  Service.

4
LBE description

• Congestion on an interface due to LBE
• Should be transparent to BE or higher classes of
  services.
• No BE or higher classes classes of services
  packet losses.
• No end-to-end guarantees
• no metric needed to quantitatively describe the
  service
• Can be supported on one interface
• Anywhere else, the LBE tagging should be passed
  transparently.

5
Application scenarios

• Mirroring
• replica update can be done at any time of the day
• Test traffic
• differentiation between the production traffic
  and the traffic generated by test of new
  applications/middleware
• Support of new protocol
• TCP extension for large data transfer
• Traffic management to from student dormitory
• research traffic protection
• Access capacity management

6
Pan-European Research Network Backbone
Topology December 2002

• Connecting 32
• European Countries
• and 28 NRENs
• Backbone capacity in
• the range of
• 34Mb/s-10Gb/s

7
Queuing technique

• For algorithm with bandwidth shared assignment,
  as Weighted Wound Robin and Weighted Fair
  Queuing, a very small bandwidth share is
  allocated to the LBE queue.
• Example in the GÉANT network (Juniper M series)

8
Queuing technique
WRR - Juniper M-series

• Weight
• Assure the queue to be given a minimum amount of
  bandwidth proportional to the weight.
• Priority
• queue with high priority are served before the
  low priority
• allow the BE (and other high priority queues) to
  be served first until empty before serving the
  LBE one.

9
Experimental results on GÉANT

• Tests were performed in August 2002.
• GÉANT - DataGRID collaboration.
• Co-existence of three classes of traffic
• Best Effort, Less than Best Effort and Premium
  IP.
• Analyse their performance in terms of
• packet loss, throughput, one-way delay and
  instantaneous packet delay variation (IPDV).
• Traffic generation
• UDP packets with SmartBits 600 - STM-16 cards.
• TCP with Netperf

10
Test infrastructure
STM-16
11
Measurement without congestion

• Packet loss
• Various LBE rate and packet size.
• No packet loss seen for LBE, BE and Premium.
• One-way delay
• Delay stable for both BE, LBE and Premium IP at
  various rate.
• IPDV
• almost negligible for the services (11µs)
• BE and LBE same behavior.

12
Measurement with congestion

• Packet loss
• No packet loss has ever been experienced for BE
  and Premium, independently of the LBE load, BE
  load being lower than link capacity.
• LBE packet loss increase with the
  over-utilisation.

13
Measurement with congestion

• One-way delay
• Increase of LBE maximum one-way delay of 1.5ms
• Increase of BE maximum one-way delay of 400µs

14
ER2002 Demo - VLBI - dataGRID

• Normal Traffic
• Radio Astronomy Data
• Less Than Best Effort
• 2.0 Gbit/s

• Normal Traffic
• Less Than Best Effort
• 2.0 Gbit/s

• Normal Traffic
• Radio Astronomy Data
• 500 Mbit/s

• Normal Traffic

15
Acknowledgements

• Spirent for the use of SmartBits 600s
• Juniper for the technical support
• TF-NGN LBE interest group
• DataGrid Work Package 7