The figuring is in an entirely new scheduling scheme and 5.8.8 needing its fair bit of time to find it's ground... 1 week / 2 weeks. It still has a problem with the highly fluctuating lengths of FAAH, but that is not BOINC's fault.

The P4 machine running the full mix except FAAH runs a perfect trickle... 13 -15 in buffer... and mostly 1 out 1 to keep the supply back-filled. The C2D with FAAH only is wonky and still shows this 10 in 10 out. Patience is called for.

Added: Someone pointed out, that the larger the buffer setting, the larger the safety factor BOINC assumes to prevent over committing. Try 2.5 or 3 days for a while. Be aware that the WCG safety is 30 work units per core per day on a single machine as at this writing.

2 days later. No downloads. I have only 2 days worth of work units left (that's OK) , but have set the buffer at 6 days. Just reporting.

Continue to have the phenomena with either FAAH in an All-Projects profile or FAAH as sole project. Given the wide spread of processing times from 3:45 hours to 8:15 hours within a single batch of 10 FAAH fetched, speculate that BOINC (now 5.8.8 for a week), does get severely confused. The 10 in 10 out has now been going on for 5 cycles. Think to have read that Dr. Anderson was applied another tweak for 5.8.9...just out http://boinc.berkeley.edu/dl/boinc_5.8.9_windows_intelx86.exe (not a recommended version, just for testing).

Well, this won't work in v5.8.xx, since v5.8.xx uses a new formula to check if a wu is in deadline-trouble or not:

Computational-deadline = report deadline - (cache-size + 1 day + "switch between projects every N hours")

Due to the formula, if uses default 1-hour switch, you-ll have this table:

Cache-size set by user - actual cache-size in project with 7-days deadline:
0.1 days (BOINC default) - 0.1 days.
0.3 days (WCG default) - 0.3 days.
1 day - 1 day
2 days - 2 days
2.5 days - 2.5 days
3 days - 2.96 days
3.5 days - 2.46 days
4 days - 1.96 days
5 days - 0.96 days
6 days - zero days.

To have a little extra safety, uses 90% of computational deadline, meaning in practice the peak cache-size is 2.82 days.

If atleast 1 wu in a project is in deadline-trouble, all work-requests to this project is blocked, until not in deadline-trouble any longer, or atleast 1 cpu is idle.

If you're attached to other projects they can continue giving you more work, but a project with 7-days deadline won't give you work as long as your cache-setting is 6 days, except if out of work from project or idle cpu.