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

Tuning
Expand Configuration and Management Manual — 529522-013
19 - 19
Multi-CPU Paths
In reference to \A, \B is a neighbor and \C and \D are non-neighbors. When \A makes a
connection to \C, the load is not distributed over different paths, but only one path is
used for all traffic to \C.
The \B records which line handler is used for \A's connection to \C to make sure that
the correct path is used. This is done by setting an entry in the reverse pairing table so
\B knows which line handler to send the packets from \C to \A.
The reverse pairing table on \B can be displayed with the SCF command:
-> INFO PROCESS $NCP, RPT \A
This displays which line handlers \B and \A are using for the connections which \A has
that go through \B.
When \A makes a connection to \D, a different line handler might be selected as the
one to carry the traffic from \A to \D. This way, the load to different non-neighbors can
be distributed among the different paths in the superpath. However, the traffic to a
single non-neighbor only uses one of the paths in the superpath.
Superpath Rebalancing
Superpath rebalancing is run periodically to correct path selection as traffic patterns
change. It has three goals:
•
CPU Matching: Make sure all source/destination pairs are using a path with the
most CPU matches available (same local/remote CPU).
•
Load Factor Balancing: Try to make the load factors (LF = ETF / TF) of all paths
within 0.5 of each other.
•
Pair Count Balancing: Spread those pairs whose traffic have no adverse impact on
load factors (LFs) over all paths in inverse proportion to their effective time factors
(ETFs).
The three goals are handled in three sep
arate steps.
1. First, CPU matching is done for each source/destination p
air by looking for line
handlers that have better CPU matches than their current owner. If more than one
path has the best match, choose the one that yields the lowest predicted load-
factor spread. The pair is moved without regard for anti-thrashing bits (see below)
or possible increase in the load-factor spread.
2. Next, the load factors are balanced. The load-factor spread is the highest load
factor minus the lowest load factor; this step tries to minimize the load factor
spread until it is less than 0.5. T
o do this, calculate the sensitivity of each path's
load factor to its total traffic, assuming a linear relationship between average ETF
and total traffic. This is used to predict the effect on the load factors of moving
traffic from one line handler to another.
Then consider moving each pair from each other line handler to the one with the
lowest load factor, and of moving each pair from the line handler with the highest