Expand Configuration and Management Manual (H06.21+, J06.10+)

Tuning
Expand Configuration and Management Manual — 529522-013
19 - 16
Multi-CPU Paths
Load Balancing
When a multi-CPU path initially assigns paths to each pair of endpoints, the traffic
pattern is usually not yet known. Load balancing is used to correct this problem as
more information is gathered by moving Expand line-handler process pairs from more
heavily loaded paths to more lightly loaded paths within the multi-CPU path. A slight
disruption occurs in message transfer occurs when pairs are changed. This disruption
is similar to what can occur when a better route is found in the Expand network and
connections are reestablished over the new best-path route.
You can schedule load balancing to occur automatically at periodic intervals or you can
initiate it manually. Exactly when you should rebalance a multi-CPU path depends on
the volatility of the traffic pattern. For example:
•
If the pattern is nearly constant, then load balancing can be initiated after a change
in the status of the multi-CPU path.
•
If the pattern changes somewhat during the day, but slowly from day to day, then
load balancing should be done after a day during off-peak hours.
•
If the pattern changes radically, load balancing should be done an hour or so into
each new traffic pattern to establish new path assignments.
A maximum of 16 moves can be put on the output change list. All the above stop when
that count is reached. Pairs on the change list are flagged with an anti-thrashing bit;
selection of those pairs for moving is avoided during the next one rebalance.
Because rebalancing is slightly disruptive, $NCP changes Expand line-handler process
pairs only at the these times:
•
When a new path comes up. (This is similar to what happens in normal paths when
a new path that has a lower TF is discovered.)
•
At configurable times during the day. You can use the SCF ALTER PROCESS,
AUTOREBALANCE command to specify when rebalancing should occur. Both the
time of day and the interval between rebalance attempts can be specified, allowing
you to schedule a rebalance when traffic is minimal.
•
Immediately. You can use the SCF ACTIVATE PROCESS command to cause an
immediate rebalance.
•
When a path goes down. (In this case, the rebalancing algorithm is not actually
used; instead, new connections are set up according to the current load.)
•
If a path is revived after being down for a defined amount of time.