[bmwg] data center benchmarking - request for further feedback on drafts

["Lucien Avramov (lavramov)" <[e-mail]>]

Hi BMWG!

We have been talking about the Data Center benchmarking draft for 
sometime. As authors, we would like to solicit more feedback. To date 
the main conversations we had were around jitter and definition on the 
first draft.

The first is draft on definition is at the following URL: 
http://tools.ietf.org/html/draft-dcbench-def-01

Regarding the first draft where we received the most comments so far, we 
would like to hear from you regarding section 4,5,6 and 7. As we have 
build these while talking to customers, switch vendors and traffic 
generator folks, we want to see if there are any other comments around it:

  4 Physical Layer Calibration . . . . . . . . . . . . . . . . . . .  6
  5 Line rate  . . . . . . . . . . . . . . . . . . . . . . . . . . .  7
  6  Buffering . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
  7 Application Throughput: Data Center Goodput. . . . . . . . . . . 13


The second draft is on methodology and can be found here: 
http://tools.ietf.org/html/draft-bmwg-dcbench-methodology-02

We would like to see more feedback especially on the second draft 
section 3,4,5 and 6:
    3. Buffering Testing . . . . . . . . . . . . . . . . . . . . . . .  7
    4  Microburst Testing . . . . . . . . . . . . . . . . . . . .  . . 10
    5. Head of Line Blocking . . . . . . . . . . . . . . . . . . . . . 11
    6. Incast Stateful and Stateless Traffic . . . . . . . . . . . . . 13


We introduce a new method to measure buffering capability of a DUT. Our 
goal with this method is to not have the end user to care about the type 
of DUT he has [cut-through / store-forward] and the test will actually 
detect and measure the DUT buffering. We then use this type of 
methodology for microburst.

Then we would like to know what you think about the head of line 
blocking evaluation. This is very important while designing data center 
networks, in order to make the proper design and deployment decisions 
based on the DUT performance. We use a generic methodology here as well 
bringing the capability to understand the impact of head of line 
blocking in a more precise way than the usual current tests which 
involve only groups of 4 ports.

Finally, section 6 is about mixing udp and tcp traffic on the DUT and 
measuring the latency for udp type of traffic while measuring the 
goodput on the tcp type of traffic.

We consolidate the feedback received so far and want to present it at
IETF90 during our BMWG meeting.

Thank you,
Jacob and Lucien