[LON-CAPA-admin] Server Configuration Questions

[Todd Ruskell]

All,

We're getting some new hardware, and that has raised a couple questions 
I'd like to ask the community:

1) 32-bit or 64-bit build?  Other than simply being able to access more 
  memory per core, does that bring real performance advantages in 
LON-CAPA?  Or are there other real performance advantages to moving to a 
64-bit build?  Or is it more like, "hey, it's 64-bit hardware, so why 
not?"  We are currently 32-bit, so moving to 64-bit will require some 
file conversion, but that shouldn't be too bad.

2) As some of you know, our IT folks are moving to a virtualized 
infrastructure.  We currently have 1 library server and 2 access 
servers, each of which has dual-core hyperthreaded hardware.  Although 
the servers are virtualized, the above hardware is dedicated to only the 
designated LON-CAPA server.

The new hardware consists of blade servers with *lots* of cores, and 
even more memory.  I've already been informed that we can't claim an 
entire blade each for the library and access servers :( but I can see 
that would probably be overkill.

So finally the question:  If you had the option, which of the following 
hypothetical situations is "better" for optimal performance?    Would 
you prefer 1 monster virtual library server with 6 cores and 12 GB of 
RAM (or more if 64-bit).  Or would you prefer 1 library and 2 access 
servers, each with 2 cores and 4 GB of RAM (all virtual, most if not all 
of the time running on the same blade)?  Or is there some other optimal 
distribution of cores and memory?  In either case, the contents of /home 
are located on a SAN.

I do know that if we went to the single-server model we'd no longer be 
able to offer up our access server as spare machine, and we'd also not 
be able to offload users to another spare machine.  Are there other 
reasons to stay away from a single "monster" machine?

Back in the day when we were really limited to the number of cores per 
physical machine, having more than one machine made sense.  But now it 
seems like having three virtual servers running on the same hardware is 
just adding an extra, unnecessary layer that slows things down.

I'd appreciate any thoughts on implementation with this virtual machine 
mindset.

Thanks,

Todd