I’ve got qnx 4.25 running on a 4 node cluster. occasionally the nodes
seems to go cpu bound, i.e running sysmon (or psinfo calls) reflects
no idle cycle. rebooting any node fixes the problem (on that node).
This behavior occurs after weeks of uptime, and seems to be associated
with UTIME having high values:
$ sin -n3 -Pidle ti
yields
SID PID PROGRAM PRI START TIME UTIME STIME CUTIME CSTIME
0 8 idle 0r — – --:-- 4294967 265 0.000 0.000
I’d be willing to bet that there is idle time, but due to an overflow,
it’s not reported properly.
What I’d like to know is whether this can impacts scheduling in any way.
IOW, is there no APPARENT idle, or is there no REAL idle?.
I’ve got qnx 4.25 running on a 4 node cluster. occasionally the nodes
seems to go cpu bound, i.e running sysmon (or psinfo calls) reflects
no idle cycle. rebooting any node fixes the problem (on that node).
This behavior occurs after weeks of uptime, and seems to be associated
with UTIME having high values:
$ sin -n3 -Pidle ti
yields
SID PID PROGRAM PRI START TIME UTIME STIME CUTIME CSTIME
0 8 idle 0r — – --:-- 4294967 265 0.000 0.000
I’d be willing to bet that there is idle time, but due to an overflow,
it’s not reported properly.
Correct on the first, wrong on the second. The idle time is in fact kept
correctly (64bit field) internally, but the psinfo interface only has a 32
bit field for the time, and as such will appear to no longer count after 4g
(~49.7 days).
What I’d like to know is whether this can impacts scheduling in any way.
IOW, is there no APPARENT idle, or is there no REAL idle?.
Since the interface doesn’t allow for extra info, sysmon etc can report
correctly, but idle is in fact running. The kernel would panic if there
wasn’t an idle process ready. That said, Rob has made a work around by
using a userland ‘idle’, checkout his software page on parse.com for QNX4.
panic if there wasn’t an idle process ready. That said, Rob has made
a work around by using a userland ‘idle’, checkout his software page
on parse.com for QNX4.