Thanks skgiven for your response.

I did not have start_delay in the config file, though it's there now. Since as I read it, it affects only client startup, it would not matter afterward?. With more VM I forced a CEP2 task to start by suspending another, and it took about 20 seconds and again it caught another task (C4CW) in the task exited, may need to restart message -- but this time only one of the other two running.

This host is limited to one CEP2 task. In practice it allows two because XP is using 3 cores for BOINC and I am running andLinux on the 4th, and it is running CEP2 also. This behavior however happens even if andLinux is shut down.

Haven't time to pursue this systematically right now but will come back to it in a couple weeks.

OK I'm back on this again, sporadically available myself but the behavior is consistent whenever I check. I still see the 2 other cores' tasks restarting whenever I start a CEP task.

This time, I wanted to force a CEP task to run. I did this:
suspended all non-running tasks except for CEP.
suspended computing altogether
suspended the tasks that had been running (both DDDT-2).
resumed computing
waited for CEP to start.
Once it was running, I resumed the other tasks.
They still instantly did the "exited with no finish file" thing and restarted.
This happens only when a CEP job starts.
Page file is nowhere close to being full now so I'm clueless what's doing this. The host becomes molasses for a good minute after a CEP task starts, until a while after it actually shows as Running.
This is a Q9450 on XP Pro, boinc 6.10.58. LAIM is on.

[Still] Can't say to observe this bogarding on neither my Windows 7 - 32 duo nor Linux 64 quad. LAIM is on of course as per the tech's advise. This is with even DDDT2 C Types on the side. Per alpha client 6.12.14 these babies grab 750MB ram each, same as CEP2... heavy models I'd say. See http://www.worldcommunitygrid.org/forums/wcg/viewthread_thread,26026#315424

One thing not understood is, that if BOINC Tasks total memory use is exceeding the setting idle/work, it would pause one or more tasks and go into "waiting for memory". Not having observed this, is the swap file big enough and the minimum size set? Mine is minimum 1.5 times RAM and unlimited growth. One would expect for the full automatic managing of VM to be quick enough to expand to meet ad hoc requirements, but who knows. At least, my VMs are in a separate partition for best performance and avoid fragmentation (Windows issue, not a Linux issue).

And yes, the CEP2 jobs start with building the model taking time. Here on Linux it takes about 15-17 seconds, on the duo Centrino near a minute (1.5Gb RAM allowed), but since I've used ThreadmasterGUI to only restrict the process to 90% per core (45 is the setting for a duo to get 90), it barely noticeble. The true hitter is when the 3rd checkpoint is stored. Though not noticing when, it does take sometimes 10 minutes before the 4th (job #3) is under way. That 10 minutes for reference is out of the 15-17 minutes I see on the Linux box for the whole task as Elapsed-CPU time differential.

--//--

NB: PageDefrag used to work on XP and Vista, but refuses to run on W7, which is when I started putting the VM in it's own partition. BOINC data_dir is also in it's own btw.

The only reason I suspended computing altogether was to suspend the DDDT2 running tasks before the CEP2 tried to start; was hoping to keep the DDDT-2 from exiting at that time. 4 GB on the host, with XP that means about 3.25 usable. Page file is huge and the raid 1 is healthy. Bumped up the memory and page file settings for boinc again and will see. Thanks

No change. How are the DDDT2 tasks exiting when they're suspended?

Well the tasks are exiting, according to messages, and LAIM has been set the entire many years I've run BOINC. I assumed the "exited with no finish file" basically meant the wu crashed and had to be restarted. The whole rigmarole of suspending computing, then suspending everything but the CEP2 task to be started, then resuming computing after CEP2 was well started, was my attempt to keep the DDDT-2 taks from exiting, but it didn't work.

Even though all other tasks were suspended, when the CEP2 task started the 3 DDDT-2 tasks in memory all did the exited-with-zero-status-but-no-finished-file thing, before I even tried to resume them.

I should say that when I'm running CEP2 only the box can run 2-3 tasks, better with 2. It handles 3-4 DDDT-2 just fine, when running onlytha. It's just that any DDDT-2 tasks get crashed whenever a CEP2 task starts. And CEP2 takes a good 30-60 seconds to show as running in the client. Otherwise I've seen no perf issues with this system.
I can tweak the VM settings more, but larger did seem to help. What I meant last message was that I bumped the % BOINC is allowed to use of the PF, to no avail.
No other drive to use. The raid 1 probably doesn't help, performance wise, but I was after recovery reliability more than performance in that case.
Appreciate everyone's thoughts. Still scratching my head.