MCF-sessionUtrecht-28-Mar-2007

From GEANT2-JRA1 Wiki

L1/L2

• Performance Metrics (PM) for different p2p services have been identified - Review of Services in Hybrid Networks (paper submitted in IARIA conference).
• Work on how to combine PMs - Add “apples with oranges”
• Now the review process starts with JRA3 and JRA4; deliverable writing starts after the Easter holidays.

Dashboard

• Requirement – Description document has been finalised. Two windows are described: KPI and Metrics ranking.
  • KPI: metrics representing the whole network that you have. Aggregation in space is used to define a metric for the whole network.
  • Metrics ranking: normal detection of values. The highest values are listed, including the most significant changes over time. The goal is to have an idea if you need to change the routing, or upgrade links.
• The layout of the GUI has already started; implementation will be completed in Y3. Deliverable writing starts after the Easter holidays.
• Remark: the inputs provided to the GUI, are closely related to the Alert System (morning presentation).

L3: Pathload

Some further tests have been done, corresponding to the remarks of the last conf call. They are presented in .

Slide 6,8-9:

• The low accuracy to measure high bitrates (1 Gbps to 800 Mbps) is caused by the short duration of the measurement probes (called packet trains), combined with the inaccuracy of a normal computer.
  • We could try to increase the duration of the packet trains (e.g. with jumbo frames).
  • You need nanosecond accuracy on the machine (while we have microsecond).
• Loop: probably caused by internal variables of the tool (e.g. from 12 to 12 and not back).

Slide 11: The number of packets inserted in the network is small, depending on the number of iterations. Pathload injects intrusive traffic for a short time (e.g. 3ms), and then lays off before the next iteration. In general about 12 MB of data is inserted during a test (about 1.5 Mbps extra data). Iperf is probing for some seconds, generating much more data.

Slide 12-16: Influence of burstiness: tests have been done with different packet sizes, and increasing burst sizes. See last slides. If the off-period is in the same order of the packet train, Pathload becomes unstable. In core network / access links: smoothing through multiplexing is expected, so this should not be a problem.

General remarks:

• What is the possible accuracy on 10 Gbps links?
  • We should have a 100 nsec resolution at least, and the load should be about 30-40% to be very accurate. But e.g. Greece has a 2.5 link with only 4 to 25% utilization.
  • As we don’t have 10 Gbps measurement nodes, we cannot measure the 10 Gbps links.
• Considering 1 Gbps: it still could be an alternative for BWCTL, which can be run anytime.
  • AH: we should go to the operational people to check the desirability. Now, BWCTL is only used during night to check if the links are ok.
  • Pathload might be “available to users”, BWCTL not.
• Possible enhancements of the measurements:
  • improve Pathload software (it’s open source)
  • try and make a kernel with a higher accuracy (nsec-kernel?) (GPS clock?)
  • work with superframes, and increase the packet train duration like this.

Further work: Test Pathload over a real network. Installation could be done on the Hades boxes (contact Jochen or Stephan). Iperf is scheduled on the SA3 boxes (Geant2 boxes), based on cron the Pathload tests could be scheduled around the same time. The Geant2 box in Thessaloniki can be used (contact Thanassis).