The smaller chunk size doesn't really matter because the TCP connection is kept alive the whole time. So what is really happening is that after 100KB have been requested HTTP headers are exchanged (a few hundred bytes) and then the next 100KB are uploaded. If the uploading client has an efficient HTTP1.1 implementation, the time spent for exchanging HTTP headers is neglegible and there is no noticeable loss in speed (maybe 1% at most).
Gnutella clients with slow HTTP1.1 implementations could cause a noticeable drop in the upload through-put. That is something LimeWire was willing to accept for a cleaner implementation of swarming.
__________________ Morgens ess ich Cornflakes und abends ess ich Brot
Und wenn ich lang genug gelebt hab, dann sterb ich und bin tot
--Fischmob |