It does look like the NM code will select APs over mesh... I bet this plays havoc with IP changing between link local addresses and DHCP addresses.
Did you expect over half of the packets in your data file to be broadcasts? Specifically 11754 out of 21587 packets were sent to the broadcast address. Bill -----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Polychronis Ypodimatopoulos Sent: Friday, May 09, 2008 3:35 PM To: Michael Stone; OLPC Development Subject: Re: [sugar] 65-node simple mesh test (and counting... ;-) Michael Stone wrote: > Data Questions: > > * Are the measurements used to make the display of 'distributions of > profile arrival rate vs. time' produced from timestamps of profile > arrival as recorded by all the laptops or by some smaller set of > 'sentinels'? > All XOs got synced clocks (by means of a broadcast packet that sets the time on all machines, providing a clock skew in the neighborhood of a single second). Then, timestamps of profile arrivals were collected from all 65 nodes, each reporting arrivals for the other 64 nodes. > * What was the general nature of the connectivity graph of the 65 > laptops? Did it change over time? > The nodes are lying in the Garden area of 1CC, and they were consistently 1-hop away from all other nodes (full mesh network) all the time. > * Did you take any measurements of background network traffic? > No, but I should have. However, we can all agree that 1CC is relatively noisy environment. > * Do you have any new insight into how the presence of NM (or of > software on top of it that depended on it) was killing your > interfaces? > My understanding is that the NM is trying its best to make ends meet in terms of what the user needs (connect to an AP/XS/mesh ?) and what connection is most "reliable" (I assume that an AP is considered more "reliable" than the mesh, but I honestly doubt it!). As a result, the NM may occasionally make the bold move to move from one connection type to another, breaking all existing "connections" (quoted because there are no TCP-like connections in my experiments), leaving stale information at various points of the software stack (eg. mesh view). > At any rate, thanks for this good work! > > Michael > > P.S. - When you produce measurements like these, please include links > to the raw data. > The raw capture is here: http://lyme.media.mit.edu/cerebro/capture-1 Pol -- Polychronis Ypodimatopoulos Graduate student Viral Communications MIT Media Lab Tel: +1 (617) 459-6058 http://www.mit.edu/~ypod/ _______________________________________________ Devel mailing list [email protected] http://lists.laptop.org/listinfo/devel _______________________________________________ Devel mailing list [email protected] http://lists.laptop.org/listinfo/devel
