
26 Cray T3E User’s Guide
Source Node
Destination Node
1
2
3
+Y
-Y
-Z
+Z
-X
+X
Figure 3.4: A routing example through the 3D torus network of the T3E.
Addressing of remote memory is managed by the External Register Set,
or E-registers. Latency hiding and synchronization are integrated in 512
+ 128 off-chip memory-mapped E-registers. The E-registers consist of
a general set of 512 external registers that manage asynchronous data
transfer between nodes by providing the destinations and/or sources of
all remote references. The second set of 128 registers are reserved for
the operating system. E-registers are used by predefined op-codes.
Each PE has 32 Barrier/Eureka Synchronization Units (BESUs), used to
implement barrier and eureka type synchronization and atomic opera-
tions. Barriers may be used, among other things, to execute SIMD codes
efficiently. Eureka operations can be used to indicate, for example, that
one PE has found a solution.
The virtual synchronization networks have higher priority for the phys-
ical channel between nodes than any other traffic. Therefore the global
synchronization is very efficient.
Comentarios a estos manuales