IBM System x3850 X5 Implementation Manual page 243

Memory Installed
Memory is installed on memory cards
1, 2, 7, and 8.
Memory is installed on memory cards
1, 2, 7, and 8 and MAX5.
Processors 2 and 3 are added to the existing processors 1 and 4
Memory is installed only on memory
cards 1, 2, 7, and 8. Minimum of two
DIMMs on memory card 1.
Memory is installed on all memory
cards for installed processors.
Operational characteristic
Performance improves as DIMMs are added to evenly
populate all ranks on each memory card.
All memory on memory cards 7 and 8 is not visible if
processor 4 fails.
The operational configuration for VMware requires that
the memory cards of the installed processor have the
exact same total memory installed. For ease of
maintenance, the best practice is to have an identical
memory configuration on each of the installed
processor's memory cards.
Performance improves as DIMMs are added to evenly
populate all ranks on each memory controller for local
memory.
Performance improves significantly when more active
memory calls use local memory on the system board
instead of the memory that is located in the MAX5.
In the rare instance of a memory card failure, VMware
operational requirements can be satisfied if all of the
memory is installed in the MAX5 or each processor has
the same amount of memory disabled. Memory access
is significantly slower if all of the memory is in the MAX5.
Performance improves as DIMMs are added to evenly
populate all ranks on each memory controller.
Processing threads assigned to processors without
memory cards have a significant drop in performance for
memory-intensive tasks.
This configuration is not an operational configuration for
operating systems, such as VMware.
Performance improves as DIMMs are added to evenly
populate all ranks on each memory card.
Memory that is installed on memory cards that are
associated with a failed or uninstalled processor is not
seen by the other processors.
The operational configuration for VMware requires the
memory cards of the installed processors have the exact
same total memory installed. For ease of maintenance,
the best practice is to have an identical memory
configuration on each of the installed processor's
memory cards.
Memory latency time increases by 50% for memory
access calls between processors 1 and 2 when the QPI
wrap card when the QPI wrap card that is installed
across QPI slot 1 and 2 has failed or is not installed. The
same effect occurs with memory access calls between
processors 3 and 4 when the QPI wrap card installed
across QPI slot 3 and 4 has failed or is not installed.
Failure of a QPI wrap card will be represented in the
hardware event log as a loss of a redundant QPI lane for
two of the processors.
Chapter 6. IBM System x3850 X5 and x3950 X5
223