[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Feedback on Texas A&M Webcast 11/20/00

Tom -

We've been doing some more testing with RN and the webcasts we do at
Berkeley.  I've just come to a big breakthrough in understanding what is
the likely problem with these live webcasts.  Put simply, the real
server has a fixed size buffer of coded data that it is streaming out to
N clients.  Each client has a variable delay due to delivery congestion
- i.e., the client buffers a certain amount of data, typically 5-10
seconds, to smooth out network congestion.  As data gets late, the
client adjusts the delay to be longer and longer up to some bound. Now
consider all the clients. Suppose the delay in a particular client is
such that when they request a lost packet, the packet is no longer in
the server buffer.  Well, the only response is to drop the stream.  You
wait a second, reconnect and everything works fine for a short period of
time again.

This fits what we saw with some, but not all, of the problems we have
been encountering both with iBeam and at Berkeley. This doesn't explain
all the problems - here are the remaining ones on our list:

1. "buffering to death" - start player and it goes into buffering state
and never exits.
2. file not available on reconnect after dropped connection 

This contrasts with on-demand playback which works very well.  The
reason is clear - you have a single buffer on the server for each client
and the source data can be read as needed from the file system.

iBeam, and all the other CDNs, want to build out edge servers with
predictable quality to the desktop.  Problem with that is two fold:

1. edge servers don't exist where the customers are located
2. even with the edge server, you may still run into the congestion
problem between the edge and the client desktop.

Interesting problems for the future!
	Larry
-- 
Professor Lawrence A. Rowe          Internet:  Rowe@BMRC.Berkeley.EDU
Computer Science Division - EECS       Phone: 510-642-5117
University of California, Berkeley       Fax: 510-642-5615
Berkeley, CA 94720-1776            URL: http://bmrc.berkeley.edu/~larry