RTFM [Read the friendly manual], an art few master these days. ;)

1) Set config to allow multiple clients
2) Learn new tricks in the RPC and port assignment department. Default is 31416, and usance is then to use the next higher up for extra clients.
3) The tricky part, you get default multiple clients with same name, so you want to play with using the device ID numbers instead.

Mind you, this is what I've read over the years, but not been practicing (very low interest), so it again is you who needs to develop the magic formula.

Sorry, what you call magic is rather a mundane thing to me... Not touched the configs in weeks on my 2x8 core clients, always crunching 2 & 1 CEP2 [work machine 1], and the rest, what I prefer. Not once had a shortfall in November, particularly CEP2 [that my alz-lite memory can pull back into the frontal lobe]. I just look in my hunting tool main page once in a while which host got what, with current mean runtimes and cahce projection, all pre-chewed for me.

One thing I have done with success while running only one other project with cep2 is setting max concurrent for both projects to 2 as in the following example:

<app_config>
<app>
<name>cep2</name>
<max_concurrent>2</max_concurrent>
</app>
<app>
<name>oet1</name>
<max_concurrent>2</max_concurrent>
</app>
</app_config>

This way the cep2 WUs don't have to be at the top of the queue. But you need to always have at least 2 WUs of both projects in IP or cores will run dry.

I think that is the correct answer on WCG. I used to devote entire machines to CEP2, but now with the limited work that is not possible. So I just let several types run at once, and let WCG set the priorities. It works, which is rare in this business.

Not sure if we are talking about the same thing, SekeRob*. I don't want to have my CPU busy with 2 CEPs more than 99% of the time. And just to make it even clearer, I was not talking about crunching efficiency either. All I would wish is to have my CPU busy with 2 CEPs ~99% of the time while also crunching other projects - meaning more than 1, as deltavee had suggested and as I have already experimented.

I hope this makes sense.

Oh, and as an update to my ongoing experiment, it's really a hit and miss thing, but more of a "miss". Indeed, tuning the overall cache is worth nothing if the WUs estimates change. I already had 2 idle cores because my machine downloaded all the CEPs it could, thus preventing other projects to reach my client.

Unless the user is ready for micromanaging, it's really not worth it. And even if he/she is, then that defeats the purpose of this whole thread.