Christopher Browne <cbbrowne@acm.org> writes:
> Centuries ago, Nostradamus foresaw when systemguards@yahoo.com (Jaime Casanova) would write:
>> Can anyone tell me if postgresql has problems with xeon processors?
>>
>> If so, there is any fix or project of fix it?
>
> Well, there's a known issue that IA-32 systems with more than 4GB of
> memory use an extension known as "PAE" to bank-switch between the
> banks of memory.
AIUI, it's not really "bank switching" in the old 8-bit sense.
Rather, there is a big linear 36-bit physical address space, and the
processor's page tables have been extended so they can point to any
page in this space. Any one process still sees a 4GB (32-bit) address
space since that's how big the registers are.
> Any time you switch banks, there's a fair little bit of work to be
> done. That includes multitasking systems that need to context switch
> a few thousand times per second.
I don't think this is any more overhead than a "normal" context
switch--cache misses, TLB flush etc.
> The "fix" for this problem is to rewrite all of your applications so
> that they become conscious of which bits of memory they're using so
> they can tune their own behaviour. This, of course, requires
> discarding useful notions such as "virtual memory" that are _assumed_
> by most modern operating systems.
This is only if you need to address more than 32-bits worth of address
in a single process, which is not always the case on server-class
systems (on scientific/numerical workloads, it's often a big win). In
which case you are certainly right:
> The fix is to buy hardware that hasn't been hacked up so badly.
64-bit systems get cheaper all the time... :)
-Doug
