I wrote:
> Cutting his value for shared_buffers (currently about 800MB) might be
> wise too. I'm not sure what the effectively available address space
> for a win32 process is, but if there's any inefficiency in the way
> the address space is laid out, those numbers could be enough to be
> trouble.
Actually, a bit of googling turns up this:
http://msdn.microsoft.com/en-us/library/aa366778(VS.85).aspx
which says that the available userspace address range for a win32
process is only *two* gig (although you can get to three using tricks
that I doubt are in his PG build). Take 800M+500M off the top, and it's
not too hard to credit that it might be tricky to swing a cat in the
remainder; especially given that it sounds like he's running complex
queries that could want to eat a lot of working RAM themselves.
IOW, these numbers are too big for your system.
regards, tom lane