NonStop Systems Introduction

NonStop Server Architecture
NonStop Systems Introduction—527825-001
7-5
Dual Access Paths for I/O
duplication of ServerNet fabrics enables processes to carry out their work by means of
a continuous exchange of messages with other processes in the system.
The ServerNet interface, shown in each processor, serves to convert messages
coming from the message system into chains of ServerNet packets or, conversely, to
convert chains of ServerNet packets into messages for the message system.
ServerNet packets are the form in which messages travel from source to destination
through the ServerNet fabric.
The distance that messages need to travel can vary considerably, and likewise the
number of routers that need to be traversed can vary. For example, a message going
from processor 0 to processor 1 needs to go through only one router (either by way of
the X fabric or the Y fabric), whereas another message going to processor 2 needs to
go through four routers. The reason for the difference is that the two pairs of
processors are physically located in different enclosures (0 and 1 in one enclosure, 2
and 3 in another). There is a separate router for entry and exit to and from each
enclosure.
In Figure 7-4 on page 7-4, the two ServerNet fabrics have been separated in order to
show the ServerNet routers that are needed to interconnect these processors. Each
router is a six-port crossbar switch (actually an application-specific integrated circuit, or
ASIC) that can simultaneously connect any input port with any output port. The black
dots represent ports that are not used in this example configuration.
At first glance, this arrangement might seem like unnecessary data handling and that a
bus would be more direct. However, consider this: all eight of the processor
connections to the two ServerNet fabrics can be simultaneously active, either sending
or receiving a message. In the case of a bus interconnection, by way of contrast, each
bus is exclusively tied up while one transmission is in progress. The disparity becomes
even greater when there are more processors in the system; there remain only two
buses, but an ever greater number of available routers. Thus, overall throughput is far
greater for a ServerNet architecture than for a bus architecture.
Furthermore, what is just as important, the intervening routers (2, 3, 6, and 7 in this
example) provide additional ServerNet ports that allow massive expandability for both
processors and I/O. The NonStop S-Series Server Description Manual discusses the
theoretical expansion that is possible because of all the ServerNet ports that can be
interconnected. Implementing just a portion of the theoretical limits of ServerNet
expandability provides an overall bandwidth that bus architectures cannot approach.
Dual Access Paths for I/O
Applications obtain the data to be processed from I/O devices and send the results of
processing to I/O devices. The ServerNet hardware and device controllers coordinate
these data transfers between applications and devices. This arrangement is illustrated
in Figure 7-5 on page 7-6.