[chord] Re: some statistics from nyu2 data

[Michael Walfish]

| Over each second, we are generating about 48.5 packets per
| second [sends plus receives], ranging from 30 to 86 packets.
| Number of packets roughly equally divided between sends and
| receives.  It appears that we receive about twice as much data
| as we send.
| 
| The packets we send behave like this:
|     2500 pts. avg =  116.531 pm   68.662 ; rng = [24, 440]
| The packets we receive look like this:
|     2500 pts. avg =  242.966 pm  128.564 ; rng = [40, 600]

This looks really thorough (thanks!) but I can't parse it. Is it the output
of a program?

| I can try and derive more statistics if you're interested.

No, thanks. I just wanted a rough sense, which the tcpdump gave. Turns out
mit5 is now using much less bandwidth. My eyeballing is confirmed by:
 http://www.planet-lab.org/logs/scout-monitor/showfile.php?filename=ts03-09-08.html

So it seems like the keyhash_mgr_timer() sending around blocks for which it
is no longer responsible every 10 seconds was indeed the explanation for
mit5's high bandwidth consumption. Either that, or the explanation is the
pmaint object (I commented out both), but my ring shouldn't have much data
stored in the regular database, since it uses only KEYHASH blocks.

Permit me to recommend that the keyhash_mgr_timer():
(a) not enumerate over its entire .k database inside one timer callback
(b) delete the block from the .k database if it hears that the recipient
successfully stored the block. once the recipient has done so, it seems
like any failure of the recipient after that should be handled by
replication and the assumptions of independent failures.

Thanks again for sending that tcpdump trace.