IBM RS/6000 SP, Руководство по определению проблем

Страница: 123 / 316

This soft copy for use by IBM employees only.
powered. A synchronous reset is performed by the switch supervisor microcode
each time the clock source (PLL or mux) is set.
•
Power-on reset (POR)
−
Flush registers (zero all bits)
−
Self-test
−
Flush registers
−
Synchronous reset
• Synchronous test
−
Error enables off
−
Clear errors
Multiple clock tree alternatives will be available in an Eclock topology file, and
so the successive alternatives will be presented one after the other.
4.4 Node Behaviors
The SP nodes can adopt different personalities when they are being part of the
switch network.
4.4.1 Primary Node Behavior
The switch commands are issued on the Control Workstation. These commands
tell the primary fault-service daemon to carry out the command.
The primary node considers the backup primary invalid if at least one of the
following is true:
•
The primary node has received an Error/Status packet from the backup
′
s
switch port, and subsequent attempts to contact the backup fail.
•
The primary node has received a bad Error/Status packet from a scan of the
network, and subsequent attempts to contact the backup fail.
If the primary detects the backup has failed, it will automatically bring up a new
one. The new backup will be chosen such that it is as “far” away from the
primary as possible, in terms of the switch and physical node location (that is, on
a different switch). The intent is to make the probability of simultaneous primary
and backup failures as low as possible.
When a switch chip detects a condition the primary should know about, it sends
both Error/Status packets to the primary, to virtually guarantee that the primary
receives the Error/Status packet at least once. The two Error/Status packets
take routes that are disjointed, with the exception of the last hop, which links to
the primary node itself.
4.4.2 Primary Backup Node Behavior
The primary backup node has two duties. First, the primary backup node detects
the loss of the primary node and initiates the takeover processing. Additionally,
but usually the duty it is performing the most, the primary backup node functions
as any other secondary node.
The backup daemon will start a thread that will queue a request token if it has
not received a packet from the primary in M seconds, where M = 300 by default,
and it is called in the primary
′
s scan interval. When the daemon receives the
Chapter 4. The Switch
103