"Nigel Griffiths" <nag@uk.ibm.com> schrieb im Newsbeitrag
news:caefuo$juc$1@hercules.btinternet.com...
> The Hypervisor and TCE are a minimum of 256MB each (TCE can be bigger on
> large machines like the p690 with many remote I/O drawers) and are like a
> Logical Partition tax.
> The other 512MB to 756MB bit is used for the Page Tables used to manage
> virtual memory (they are virtual address to physical address translation
> tables used by the CPU H/W).
> After an initial chunk, this is 1/64th of memory.
> This can be seen in LPAR machines because the Hypervisor is managing
virtual
> memory to ensure the LPARs can never share the same physical memory (LPARs
> are isolated) abd is seen on one of the HMC machine proprty panels.
> Exactly, the same amount of memory is used when not in LPAR mode - i.e.
SMP
> mode or just running on a normal RS/6000 or pSeries machine but you are
> never told how much memory is used by these same Page Tables.
> So yes the 512MB is an overhead you pay for LPAR but the rest is not.
>
> Hope this helps, Nigel Griffiths
> >

Hi Nigel,

thanx for the theorie ;-) However the example I gave is from the real world
and the IBM customer (international consumer goods industry) who suffered
that 756 MB penalty for trying to us LPAR on his 6C4 with 4GB was really
p*ssed off because he could only use 3,25 GB per server in LPAR mode.
Actually he bought 4 servers with 4GB each and so lost 756 MB per server.
This amounts to about 3 GB of ram i.e. he lost almost one complete server's
memory. Hence he decided to use Full System Partition with two servers in
order to use the whole ram at least on those machines but at the same time
he was quite dissappointed of his IBM sales representative because that guy
did not sell him what he wanted. What can IBM learn from that? If a customer
wants a server with 4GB of RAM with an LPAR ready pSeries better make sure
that he not only gets 4GB but that he can *use* 4GB.

Regards,
Andreas