It's difficult to explain by itself, but nearly impossible to do so without you giving a clue as to what your hardware is (amount of CPUs and GPUs really), and which projects you run, at what resource share (you mentioned that in passing), switch between (science) applications time and whether you use the default (project given) applications of specially optimised ones for additions to your CPU (MMX, SSE, SSE2 etc.) Then what your setting is for the connect to interval, plus the additional days worth of cache.

Topped off with whether you run BOINC continuously 24/7, or only let it do work when the computer is idle, or that you only allow BOINC to do work for (say) 8 hours a day and shut the computer down the rest of the day.

Lastly, whether you let BOINC do all this for itself, or that you interfere (intervene) by pausing some projects to allow others to go do fetch.

But most of the time the fluctuations you may see is due to debt and projects having no work or them being down.

Tech stuff you may want to look at:
http://boinc.berkeley.edu/trac/wiki/ClientSched
http://boinc.berkeley.edu/trac/wiki/GpuSched
http://boinc.berkeley.edu/trac/wiki/GpuWorkFetch
and just to throw all those out the window, http://boinc.berkeley.edu/trac/wiki/ClientSchedOctTen (the latter being the new scheme being tested in 6.13).

---