Expand Configuration and Management Manual (G06.24+)

Tuning
Expand Configuration and Management Manual523347-008
20-15
Multi-CPU Paths
The general formula for configuring the variable packet size for multi-line paths using
low to medium bandwidth lines is as follows:
If the average message size is not known, you can use the Expand subsystem SCF
command PATH STATS to display a histogram of message sizes.
Multi-CPU Paths
The multi-CPU path is the fundamental component of the Expand multi-CPU feature. A
multi-CPU path can consist of up to 16 individual Expand paths, including multi-line
paths. Expand-over-ServerNet and Expand-over-FOX line-handler processes cannot
participate as members of a Superpath.
The main advantages of the Expand multi-CPU feature include the following:
Higher total throughput
More even spreading of the communications load over multiple processors
Reduced message system interprocessor traffic
The main disadvantage of the Expand multi-CPU feature is that its advantages are
available only when traffic fits a certain pattern. For example, if most traffic occurs
between the same two nodes—or if these nodes are direct neighbors and traffic is sent
between the same two processors in one direction—then the Expand multi-CPU
feature cannot spread the load effectively. Other disadvantages include the following:
Increased processor overhead
The possibility of occasional disruptions during load balancing
The interaction of some elements of the system and of the Expand network determine
the degree of performance improvement that the Expand multi-CPU feature might
achieve. These are the elements that control how well the Expand multi-CPU feature
can spread traffic over its constituent paths.
These elements include the following:
Traffic Pattern
Load Balancing
Traffic Pattern
Unlike a multi-line path, which can spread traffic evenly over all of its lines regardless
of the traffic pattern, a single path within a multi-CPU path is assigned to traffic
between each pair of endpoints. This path assignment is fixed until traffic is rebalanced
over all the paths in the multi-CPU path. For this reason, the more that traffic is spread
across different endpoints, the better a multi-CPU path can spread the load across its
member paths.
PATHPACKETBYTES = average_message_size / number_of_lines