I'm no expert on the server-side options of BOINC, but a search of the BOINC site shows the following seemingly relevant config options (at http://boinc.berkeley.edu/trac/wiki/ProjectOptions):

N
N

Here's what it says about that last option:

min_sendwork_interval
Minimum number of seconds to wait after sending results to a given host, before new results are sent to the same host. Helps prevent hosts with download or application problems from trashing lots of results by returning lots of error results. But don't set it to be so long that a host goes idle after completing its work, before getting new work.

What we are seeing on fast hosts, particularly with short (2 credit) work units is exactly as described - they run out of work before they are allowed to connect again... This may become even more prevalent if the new applications are faster.

Perhaps if that were set to 5 or 10 minutes, things would run more smoothly

dave set a recent response that should hopefully fix this, so i'll get labstaff to get the newest version of BOINC and update the server.