TLDR version: Symlinked, same/diff directories showed no noticeable difference in memory use whether it was 1 or 5 instances of wow. Why?
I (try to) 5-box on one Mac Pro with the following specs:

- 2 Quad Core 2.8 GHz (8 cores total)
- 4 GB ram
- nvidia geforce 8800 GT 512 MB
- 1 stock 320 GB (I think Seagate Barracuda 7200.10)
- maxfps: 30, maxfpsbk: 10
- sound off
- lowest video settings except main has max view distance
(amazing to think these are standard specs on his board these days /feels bad for his old windows machine)

Outside of cities, it works modestly well. It's rare to lose a follower (I use Jamba follow strobing - excellent stuff). I do find I have more difficulty if I turn up settings. Porting into Dalaran is a nightmare however. I almost always disconnect someone if I do all 5 at the same time (the main having a load of addons may contribute to this). /follow through Dalaran is impossible.

So I started digging around and have been using the method of symlinking the directories for a while now. Still disappointed with the results, I wondered about the expected efficacy here. To me, it makes sense that the OS would recognize the same files are being requested and only pull them from the HD once, only needing one copy, right? Well, one test led to another and I'd appreciate another perspective on these results.

How I set up the tests
I wanted to measure memory usage under running a varying number of WoW instances:
- from 1 directory, no symlinks used
- from different directories, no symlinks used (yes, 5 Data directories take a long time to duplicate)
- from different directories, all using symlinks (that is, all 5 point to a Data directory not being otherwise loaded in the test)
- from 1 directory, symlinked to another, otherwise unused directory, in Dalaran

For each of the above conditions, the machine was rebooted and no extraneous software was loaded. In the first three, the toons were standing in the same spot with the same view at the mailbox in Warsong Hold, Borean Tundra. The Dalaran test was conducted with the toons standing on the sewer grate at the entrance to the Horde area and the camera at max distance directly overhead.

What I found was that there was very little difference between conditions (except Dalaran) and little to no reduction in memory used per instance of wow as additional instances were started. At 5 instances and ~1500 MB of ram used for Wow, there was a less than 10 MB difference between setups. I also monitored what was labeled as "Virtual Memory", "Private Memory" and "Shared Memory". Again, those had differences of around 10 MB. Free memory exhibited similar behavior with a slightly larger variance. The Dalaran setup obviously used a significantly higher amount of memory in all cases.

Perhaps I lack a deeper understanding of the innards of how Mac OS reports memory info to Activity Monitor (assuming via top). There were 4 available columns related to memory and I recorded those in the spreadsheet. You can see the everything else in the screenshots of activity monitor if you're curious.

The simpler-to-read spreadsheet does a better job of lining up the relevant numbers for comparison. Each 'sheet' is a condition of running 1, 2, 3, 4 or 5 instances of WoW. Vertically on the sheet, you'll find the four differences in directory usage and the readings of memory usage for each instance as well as a total.

Reading this, I'm surprised by the lack of difference between symlinking and not. I thought (based on my readings at this board) that the symlink results would be best from and the 5-data-directory setup would have the worst memory. While I never expected 5 symlinked copies to use 1/5th of the memory, I was shocked to see it used more than 4/5ths the memory and debatably no difference. At first, I thought it was because my 5 instances pretty much maxed the memory anyway, so that's why I tried 1-5 instances. However, those provided the same results.

Can anyone explain this? I hope this sparks some solid, discussion.

* This is by no means a very scientific experiment. There were way too few samples and not enough controls. Also the person planning the tests (me) does not have extensive knowledge of the behind-the-scenes bits of memory management. So take it with a grain of salt.

Thanks for your time

What this does not show is the more realistic performance of actually running around while random textures and models are retrieved from the harddrive as I didn't have at good method for testing that. Any ideas are welcome.