cc'ing creis@ for more insights on this bug.

Based on offline Chat with creis@ "--process-per-tab" isn't really supported flag. 

creis@ can you please give more insights on what  niklas.gollenstede@ is seeing.

Based on steps 1a through 2b, it sounds like the problem is that the process limit may be counting Chrome processes from other --user-data-dir sessions (profiles).  I'll see if we can repro that locally, since my first reaction is that we should let each profile see its own process budget.  (I agree that effectively forcing a second profile into single process mode isn't a good outcome.)  CC'ing nasko@ for this, since we both know about the limit.

I think the --process-per-tab aspect is likely a red herring.  It's not meant to change the process limit-- it prevents process swaps from happening for (browser-initiated) cross-site navigations in a tab.  See:
http://www.chromium.org/developers/design-documents/process-models

As for renderer processes being limited to about 3-3.5 GB, I'll have to defer to someone else, and it should probably be tracked separately if it's not intentional.  rschoen@ or hablich@, do you know if this should be considered a bug, or who to ask?

The upper limit per process is currently working as intended. Since a few weeks ago V8 is supporting larger heaps which will enable Chromium in the future to support higher per process memory limits.

Thanks for your replies!

After having read about the process models in more detail, I'd agree that "--process-per-tab" actually has nothing to do with this.
Also, I now think that multiple-tab "related browsing contexts" may be the actual cause of this. When I wrote to "Open another bunch of tabs" in step 2b, I have always done that by Ctrl+click'ing different links from a single tab, which (probably?) causes them to be related to their origin tab - even though there is no opener relation between them (only the document.referrer string) and I don't think they can actually access each other.

So today I did step 2b without any other chrome instances open, and was able to produce the described effects without steps 1a-d.
(I was reasonably sure that steps 1a through d are actually required to cause this, but maybe I was mistaken.)

I'll have another look at this tomorrow.

@niklas.gollenstede: Gentle ping, can you please provide an update as per your above comment.

@creis:Request you to please take a look and respond as per the update provided in comment#6 if required.

Thanks.!

Closing issue as Wontfix due to lack of feedback requested but not provided. If the issue still exists please open a new issue with the details requested.

Thanks..!