I think you will find your preferences are set to use at most 50% of your memory when the computer is in use.

Change your preferences.

Also, if you have a dual core computer, you must double the minimum requirements.

Average peak memory use for The Clean Energy Project is 145 MB. Average peak page file use is 429 MB.

Having no doubt that BOINC regulates permitted idle/use ram level use correctly, my own observations on CEP jobs have shown quite a bit larger VM use than Didactylos and incidental larger RAM use. If thus the sum of the actual disk space use or memory is exceeded, 1 or more sciences will be paused and a smaller one will be tried.

Now, in client 6.6 a new logic is introduced and for each application BOINC is now recording the maximum encountered for each. This stops the client from trying a whole string of jobs to see if they will fit. The reason is that if LAIM (Leave Application in Memory) is on, ever more paused jobs would sit in memory. In past BOINC would just try job after job to get all cores active. This for me has a good/bad snag. If a client only runs the CEP project and it's all there is in queue/cache, with too little memory the client will not run those jobs concurrently with too tight memory constraints. What will happen is that as soon as the job that was still running finishes, the waiting in memory job will be started, and this is a guess, no second job will even be attempted to be started based on past historic maximum use record. I'm curious as how the developers will signal this in the client log.

Now Haris Dublis, can you please post the actual startup log from the stdoutdea.txt file from before your preference changes, for that will tell what the effective net memory/disk use is. The percentage can be misleading. It's what BOINC computes net off, reflected in these log lines:

21/04/2009 10:09:38 Preferences limit memory usage when active to 2302.32MB
21/04/2009 10:09:38 Preferences limit memory usage when idle to 2762.79MB
21/04/2009 10:09:38 Preferences limit disk usage to 10.00GB

Cant's tell which batch, just mental notes when looking at properties in the Process Explorer. Average Peak I assume is not the absolute peak which I've seen hit 490Mb and RAM near 200Mb, exceptions, but still occurring.

When I read your first post I had a look at a B type which was running under XP: it was using 210 MB RAM and the VM size was 487 MB. I forgot to look at the peak RAM.
But it was a big WU, it took 12 hours of runtime on the quad. Batch 485.

The A type (batch 511) which is running now (under Linux) is much smaller. Top says 20 MB RAM and 245 MB and it should run only a little more than 6 hours.

Cheers. Jean.

Haris Dublas writes in the plural " I guess I have to remove CEP on my systems with only 1GB to avoid this problem in the future.". Single or dual core, with 75% memory use allowed to BOINC and enough VM to go around and disk storage space permission, it should be enough to have both continue.

As also he writes that at night they do nothing else, not likely they get paged out, and even them, they'd keep running just doing more VM swapping, long as the minimum RAM threshold is not causing the waiting in memory.

One of the pains of updating the website device profile is the necessity of the client to talk to the server to fetch the new settings. If in the middle of a longer job that could take some time and with waiting for memory even longer so hitting the Update button in the BOINC Manager project tab is always advised if wanting to force the revised profile across.