bharry's WebLog

I'd have sworn that I'd written more about this before but I can't find it.  Here's a post that talks about the disk configuration we are now using: http://blogs.msdn.com/bharry/archive/2006/07/24/676594.aspx

I can't find the precise details of our installation at the moment, but I know the basic stats.  We use a "3" server configuration roughly as follows:

application tier - 2 proc, dual core with 4GB of RAM running 32 Windows 2003

data tier - 4 proc, dual core with 32GB of RAM running 64-bit Windows 2003

analysis server - I don't remember the stats on it but I think it's a modest database server much smaller than the data tier.

Brian

The reason for so much memory has to do with the size of the file set.  Any recommendations for size are inherently flawed because to make them, you have to make a lot of assumptions.  When we put out guidelines, we make the best assumptions we can about what is "typical" and base the guidelines on that.

Our installation isn't typical at all.  For one thing, we sized the hardware for 2,500 people - because that's what we expect to use it eventually.  Secondly, our projects are MUCH larger than is typical.  A single branch is now over 2,000,000 files and the average developer uses 100,000 or more of them at a time.  Further, we do large branches, merges, checkins, etc that frequently involve hundreds of thousands or millions of files.

Lastly, I just wanted to be safe when we speced out the machine.  I wanted a margin of error so that we knew memory would not be a problem.  In truth, based on what I know now, if I were publishing the recommendations today, I'd probably change the 800-2000 config to be 8GB on the data tier.  I think it would be a safer bet.

You can read a post I made a month or so ago about the version control performance improvements we've made for Orcas.  One of the big effects of these improvements is that it should reduce memory pressure of large operations both on the AppTier and DataTier.  I suspect once we get that rolled onto our dogfood server, we could drop our memory to 16GB with no problem but we probably won't since the memory is already in there.

Brian

Sorry, missed the question about Analysis Server.

I'd start with it on the same server and observe utilization.  We moved it because we have so much data in our warehouse (see my dogfood stats) that AS would run for 20 or 30 minutes rebuilding the cube every hour and we didn't want that much load on the server.

Like I say, I think the best thing to do is observe the behavior and adjust based on what you see.

Brian