The major topics discussed were the following:
Assumptions:
The discussion that ensued about the assumptions of the paper
were centered on the analyzing the premise that only large flows need to counted.
Is the accounting of only large flows providing the information needed about the
network? For the scope of the paper, and the design of the algorithms, it was
agreed that the assumptions are valid. However, a counterpoint was raised that
as a links capacity increased, small flows that used to be of interest may no
longer pass the threshold of these algorithms. So, lowering the threshold to
make up for the loss of data would require more memory and would bring down
the efficiency of the algorithms. However, this argument was countered with
the original assumptions of the paper.
Memory:
SRAM and SDRAM as the storage was another topic of discussion. Is SDRAM fast enough
for the link speeds and sampling rates? Is SRAM, while being faster then SDRAM, too
expensive? These questions were discussed with the added topics of the trade-off between
the memory sizes needed for various thresholds. Also, the topic of measurement intervals
in relation to memory sizing was discussed. The presenter was able to point out that
the associated technical report had in depth analysis of the trade-offs between the
interval and memory size.
NetFlow:
The discussion about NetFlow centered around the idea that the comparison
of the two created algorithms to NetFlow was not valid. NetFlow tries to
do statistics gathering while the others try to track only the largest flows
in the network. It was acknowledged that the design of the new algorithms
is inherently different than NetFlow.
Measurements:
For the experiments that were done, concern over the unidirectional accounting to
test the algorithms was raised. The question of memory requirements being doubled
for a real measurement occupied the conversation. It was agreed that the unidirectional
accounting of the experiment was not necessarily a bad measurement. Other topics
discussed about the measurements centered on the results reported in tables 5, 6, and 7.
Average error rates and unidentified flows rates between the algorithms brought
our the comment by Jing Lu that there is a increase in the unidentified flows by NetFlow
that is 90 times that of the Sample and Hold algorithm.
Corrections in the paper:
Jing Lu: One small mathematical error was noticed. In section 4.1.1, the relative
error is stated as T*sqrt(2-p)/p when it should be (sqrt(2-p)/p)/T.
Undiscussed topics raised in reviews but not mentioned in the discussion:
Jack Meier: Commenting on a method of accounting that was not mentioned in the paper,
Jack pointed out that from "Digital Video Process by Tekalp p. 374" offers an
optimal prediction method using linear minimum mean square error.
Shobana Padmanabhan: As an alternative to the passive measurement of the algorithms,
Shobana pointed out that "Active Measurement" might be viable. The source of her
reference is "The Measurement Manifesto by George Varghese and Cristian Estan".
Ranking the paper from the members of the seminar dominated an opinion that a middle third position was warranted. However, it must be noted that there was no one who put the paper in the bottom third ranking while a small group thought the paper warranted a top third position.