Not sure how you connect but one thing that helps when up/downloading for me is changing cc_config.xml to include the following:

<max_file_xfers>31</max_file_xfers>
Maximum number of simultaneous file transfers (default 8).
<max_file_xfers_per_project>30</max_file_xfers_per_project>
Maximum number of simultaneous file transfers per project (default 2).

You might need to adjust the actual number to suit the capability of your internet connection

Assuming that you have your profile on this machine set to Minimum Work Buffer 8 days (if you collect new work every 7 days), this should get you a shed load at once.

Randzo, how much time does it need to upload your 82 results of the previous period? The concurrent number has to my finding, little effect on the total end time... the more I have uploading at the same time to Harvard, the slower it goes. Downloading the much smaller Task files I see completing in infinitely shorter time [except when there's a project [re]launch fetch storm i.e. whether you get 3-4 per fetch or 40, at the end the fetches are probably finished before uploads complete. Yes / No?

Fact: Every time you hit the update button you reset the deferral counter, so better keep your hands off or the 182 seconds restart after a [too early... last request was only xyz seconds ago]. Monitor in the Projects tab of BOINCTasks/BOINC Manager e.g. where you can watch the counter life to avoid jumping the gun [you will probably already have found out].

Whilst, I just puzzled on something only now hitting me [again]. If there's more than 2x results to upload than BOINC permitted processor threads, download of new tasks is blocked [your ~80% upload completion swivel point]. knreed et al would have to comment on this real snag [think this is a very old rule, no matter what client version]. Don't know what his special rule / exception leeway is on giving more CEP2 work per call. My octo, until that SPO condition developed last weekend, never seemed to get more than 3-4 per call, but then the total was allowed to be 10 in the profile and networking is fairly constant, albeit slow. ~82 per fetch cycle is a special problem.