Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: [802.3AZ] about trace files



Greetings Mike,

One thing to bear in mind is that changing the rate will greatly
change the packet arrival process, because TCP will react.

I saw a paper saying "if we reduced the link rate using policy X, we'd
get huge packet backlogs" because it was simply using layer-2 packet
traces.

It would be good if your capture included enough payload to be able to
reconstruct TCP sessions, the way Tmix does.

Cheers,
Lachlan

On 22/04/2008, Mike Bennett <mjbennett@lbl.gov> wrote:
> Folks,
>
>  A number of people have been talking about the use of trace files to
> understand the nature of the link utilization for various types of networks.
>  Before I start a campaign to build a library of trace files, I'd like to
> make sure my assumptions are correct so we get a useful set of files.  Once
> I get agreement from folks on what we need, I will start collecting traces.
>  Bruce Nordman has volunteered to be the capture file "librarian".
>
>  Here are my assumptions (based on my experience and Bruce's input):
>
>    * We only need to capture enough of the frame to get the time-stamp
>      and frame size (frame capture length is 14 Bytes).
>          o We minimize security policy issue with sharing the captures
>            if we limit the content of the frame captures.
>    * Most open source capture tools support the packet capture library
>      (libpcap) so trace files should be saved in that file format.
>    * We need to balance the duration of the capture period with the
>      size of the packet capture file.  Based on some tests I've run, a
>      moderately utilized 10G link will generate approximately an 4 GB
>      capture file in 2 hours.  I believe 2 hours is long enough to tell
>      us something about the nature of the utilization of the link.  If
>      you want to characterize a particular link for a longer period
>      then the capture should be restarted to limit the file sizes.
>          o we could capture in 1 hour increments - it would cut the
>            file sizes down. Let me know if you have a preference.
>    * Meta data on the packet captures should include:
>          o Date of collection
>          o Name and company of individual who collected the trace
>          o Link speed and type, e.g. 1000BASE-T, 10BGASE-LR, etc.
>          o Type of devices at the ends of the link, e.g.,
>            router-switch, desktop-switch, switch-switch, etc.
>          o General usage, such as department file server, desktop PC in
>            office, HPC cluster node, etc.
>
>  This should be enough to help us determine how EEE will behave under the
> conditions of the capture.  It won't be perfect but it will help.  Please
> ask your questions or point out what you think is missing or needs to be
> changed by the end of this week.
>
>  Regards,
>
>  Mike
>


-- 
Lachlan Andrew  Dept of Computer Science, Caltech
1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA
Ph: +1 (626) 395-8820    Fax: +1 (626) 568-3603
http://netlab.caltech.edu/lachlan