Per task upload retry policy leads to kind of DoS attack from fast hosts in periods when project's server can't handle all incoming connections. When many results trying to resend themselves on it' own time schedule host _continually_ tries to connect to server, w/o any delays. With mny hundreds (and this number easely achievable now fro CUDA enabled hosts) of results in transfer queue there are always some that try to reach server at any given moment.
All that need to be done is to change retry policy from per task to per project basis (as it is for result reporting already).
This inefficiency in BOINC leads to increase conncetion problems with overloaded project servers.
(I setted component as Work fetch policy as closest to this component. Actually it's work upload, not fetch, policy)
For more descriptions of this issue look this thread and links in it:
http://boinc.berkeley.edu/dev/forum_thread.php?id=4080