Sebastien Roy <> wrote: > On Mon, 2009-09-21 at 12:22 -0700, Eric Saxe wrote: >> Sebastien Roy wrote: >>> As I mention in 6883663, my system is mostly idle. I resolved the >>> print manager issue a while ago on my laptop, yet my CPU is still >>> pegged at 100%. I brought this up a few weeks ago in the thread >>> entitled "pm confusion and build 122" with no resolution. >>> >> >> Sorry, in my following of the thread, I thought the X activity was >> causal here...and perhaps in your case that's only partly true... >> What % of time are you spending in C0 on the idle system? Does >> powertop still show frequent wakeups? > > It is in C0 less than 1% of the time. Here's a snapshot of the > powertop output: > > OpenSolaris PowerTOP version 1.2 > > C-states (idle power) Avg Residency P-states (frequencies) > C0 (cpu running) (0.9%) 800 Mhz 0.0% > C1 4.0ms (99.1%) 1060 Mhz 0.0% > 1330 Mhz 100.0% >
Ok, we know P-state can speed down if Xorg is disabled, right? Print manager is part of X activity, I think there must be other X activities causing the system utilization. Why event-mode p-state doesn't lower the speed, Because it captured the utilization indeed. prstat shows that Xorg has ~0.5% cpu utilization, vmstat shows ~1% cpu utilization. That's why poll-mode p-state can lower the speed, because the idle high water mark value of poll-mode is 85%, that means it can lower the speed if the cpu utiilzation < 15%. Apparantly it does not capture the desktop-print-management-applet bug. Event-mode p-state can capture this kindof mis-behavior. But the problem is, it's not be able to aware the cpu utilization is very low, especially the ping-pong cause we encountered. It looks like we still have a hole in the event-mode p-state for the ping-pong case. cpd_tw_governed and cpd_ti_governed will bypass our implementation to avoid the transient workload and transient idle. Instead of cmt utilization, It looks like the cpu utilization is the user prefered system index. The user expects the lowest cpu clock if the cpu utilization is very low, not care about if there is a looping bug like desktop-print-management-applet. This probably is the right expectation. So, do we need to enhance the event mode p-state governor to sampling based? Or do we need to set poll-mode as the default p-state governor? Thanks, -Aubrey
