Like clusters, these systems provide
scaling of memory and I/O subsystems in addition to CPUs. A key difference is the
single operating system copy that manages the entire platform and a directory-based
cache coherency scheme to keep data synchronized. Memory access between nodes
is in the hundreds of microseconds, which is much faster than going to disk in clustered
configurations, and only slightly less swift than local memory bus speeds in a
single SMP system. Memory capacities can range into multiple terabytes.
Figure 12-5. Non-Uniform Memory Access (NUMA) configuration
Memory Bus Links
Typical NUMA Node
Memory Bus Link
CPU CPU
L2
Memory
I/O
Disk
CPU CPU
L2
Memory
I/O
Disk
CPU CPU
L2
Memory
I/O
Disk
CPU CPU
L2
Memory
I/O
Disk
Grid Computing | 299
This enables NUMA to have some major advantages over cluster solutions:
??? Parallel versions of applications don??™t need to be developed or certified to run on
these machines (though additional performance gains may be realized when
such applications can be tuned for NUMA).
Pages:
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649