| The outermost level |
 |
a
general-purpose medium-grain MIMD architecture based on commodity SMP CPUs. Actually, this
section is realized by an 8-node (dual HyperSPARC ä /100 MHz)
CS2; |
|
a low-latency,
high bandwidth fat-tree interconnection network, adopting wormhole, adaptive, and virtual
channel strategies for achieving high predictability of communication costs; |
|
communication
coprocessors supporting message passing (MPI) and shared memory directly; and parallel I/O
and a parallel file system |
|
| The innermost fine-grain level |
|
a SIMD computer
structure - namely, a 128/512 node Quadrics-APE100 machine - connected to each of seven
CS2 nodes. Each SIMD "booster" supports the execution of instances of highly
parallel data types whenever the programmer knows and/or the compiler recognizes that such
a mapping is more efficient than using the SMP processor. |
|
| The communications |
|
the
communication and sharing between the SMP CPUs and the SIMD part of the nodes are
implemented by a structure (including HiPPI channels, PCI buses, communication
coprocessors and ad hoc components) that emulates the memory sharing between the sections
of the node. |
|
| The
software layers (top-down) |
| System and application
libraries |
|
mathematical
libraries; |
|
discipline
libraries (e.g. for CFD, chemistry, structural analysis, etc.); |
|
specific
packages (e.g. seismic migration); |
|
tuning tools,
debuggers, etc. |
|
| Programming
environment |
|
structured
parallel programming environment (SkIE); |
|
Fortran
programming environment (including HPF and QFOR); |
|
tools for
integrating TAO programs into HPF and SkIE programs; |
|
etc. |
|
| Run
Time Environment |
|
message passing
+ virtual shared memory |
|
| Operating
System |
|
MIMD section:
SUN-Solaris 2.5 ä -based CS2 O.S., integrated with a Parallel
File System and other typical MPP components; |
|
SIMD section:
Quadrics/APE100 O.S.. |
|
HPCN at ENEA 
Back
Home
page