Hitachi Compute Blade 2000, User Manual

Page: 1071 / 1927

1020
12
Log
ical
par
titio
ni
ng man
age
r
The memory access cost from the processor in Node0 to the memory in Node3 is larger
than the access cost to the memory in Node1, since it is necessary to go through Node1
to access memory in Node3.
Memory accesses on the NUMA configuration
(in the case of four nodes)
In the server blade, which is set to non-NUMA, memory is automatically interleaved and
any programs do not need to access the memory, the memory access performance to
all the memory area becomes middle of the memory access performance of local
memory and remote memory.
Memory accesses on the Non-NUMA configuration
Physical
memory
Physical
processor
Node0
Processor 0
Node1
Processor 1
Memory
access cost
Memory access
Memory access is
dispersed
I nterleaved
Memory of (Non-NUMA)
Interconnect
Node0 Node1
Processor 1
Memory
access cost
Memory access
Interconnect
Physical
memory
Physical
processor
Node2
Processor 2
Node3
Processor 3
For Processor 0, the memory in Node 0 is local memory, the
memory in Node1, Node2 or Node3 is remote memory.
Processor 0
small
large
large
Physical
memory
Physical
processor