The option is CEP2 specific only. If you have a 4 core PC and put that option so it chooses to run 2 CEP2 Wu's it will download 2 CEP2 wu's plus 2 other projects WU's you have selected.

There is no way to do what you want except for manually downloading extra WU. You actually answered your own question. That is the way the option works.

Server 700 has ways to instruct a client to report a type of task immediately [something for techs to check if they have not already], which then opens the door to fetch a new task of CEP2 [send by highest priority over any other project]. Then it's a question of setting the cache low [e.g. 0.01 on a client 6, or 0.0 on a client 7 **] and set the number of CEP2 tasks in the device profile to the maximum allowed concurrent [previously proposed in this thread], but in the advanced schema plus 1, 3 was the maximum without errors, so 4 once running full flow. Mix this with 1 or more other WCG projects... avoid setting the "if there is no work..." as that's a backdoor to slip anything across, including CEP2.

The procedure to kick this off:

1) Fetch 3 CEP2 and babysit, so they wont start simultaneous [critical]. If checkpoints logging is set, manually start the next after 1st checkpoint has been recorded, till 3 are running.
2) Fetch a 4th task of any other science.
3) Once running on all cores, set the allowed CEP2 to 4, so when next time work is fetched, it will be a CEP2 task ready in queue.

What is theorized to happen is when the first CEP2 finishes, there's technically still 4 on the system as maximum, 1 ready to report and 3 running. The 4th slot is used by any other science. Then when the client is ready to report and fetch new work [a combined action], a CEP2 task will be ordered first, by said feeder priority.

Little trial and error, but most of the time I have never more than the desired number running on an octo, and if it's less, it's less. The maximum concurrent is controlled and no babysitting is needed. Most importantly, no crashed tasks or running too many at poor efficiency. Then when I see less than the max running, but one or more sitting at end of buffer, start them manually... forced ahead of queue. Not optimal, but best I will do and getting a 100% valid score.

** The client 7 buffering is designed to pre-fetch a task several minutes before a job finishes when the buffer is set to 0.0.

Once upon a time there were some volunteer coders who had the chops and implemented a method so the private build client would not run more than X of a science_app concurrently [for such reasons as overload and heat protection as some run really hot or really inefficient when more than X concurrent]. Don't know what became of it... purposely forgotten maybe, but with ever more cores in systems in home sytems deserving renewed attention. Similarly there are private builds which have a DCF per science_app... don't wanna know why this cannot be adopted by the developers [considering I've seen it mention on the alpha mail list i.e. no secret].

I've recently re-started crunching for WCG. Downloaded a CEP2 workunit. Let it run for several hours. No stopping work, restarts, no nothing, no pauses even (but LAIM is checked). BOINC manager is showing 45 minutes elapsed and some 3 % done. Is that normal?

Result Name: E218945_ 874_ I.50.C39H20N4O6S.00321626.4.set1d06_ 1--
<core_client_version>7.2.33</core_client_version>
<![CDATA[
<message>
aborted by user
</message>
<stderr_txt>
INFO: No state to restore. Start from the beginning.
[12:05:32] Number of jobs = 16
[12:05:32] Starting job 0,CPU time has been restored to 0.000000.
[12:12:49] Finished Job #0
[12:12:49] Starting job 1,CPU time has been restored to 401.953125.
[12:31:38] Finished Job #1
[12:31:38] Starting job 2,CPU time has been restored to 1450.468750.
16:25:20 (4480): No heartbeat from core client for 30 sec - exiting
No heartbeat: Exiting
[16:25:26] Number of jobs = 16
[16:25:26] Starting job 2,CPU time has been restored to 1450.468750.
Abort requested: Exiting

</stderr_txt>
]]>

CEP2 is now allowed to run for 18 hours rather than just 12 hours, so if the estimated speed is moderately off, you could get these values. Let it run.

Lawrence