2.6.17.1 perfmon2 patch for review: PMU context switch
In-Reply-To: <[email protected]>
Quoted 13 lines · Stephan Eranian
Here is what I get on my dual 2.4GHz Opteron 250:
booted with idle=halt
$ pfmon --us-c -ecpu_clk_unhalted,interrupts_masked_cycles_with_interrupt_pending,interrupts_masked_cycles,cycles_no_fpu_ops_retired -k --system-wide -t 10 <session to end in 10 seconds> CPU0 11,356,210 CPU_CLK_UNHALTED CPU0 0 INTERRUPTS_MASKED_CYCLES_WITH_INTERRUPT_PENDING CPU0 3,836,107 INTERRUPTS_MASKED_CYCLES CPU0 23,910,784,532 CYCLES_NO_FPU_OPS_RETIRED CPU1 19,303,632 CPU_CLK_UNHALTED CPU1 0 INTERRUPTS_MASKED_CYCLES_WITH_INTERRUPT_PENDING CPU1 13,942,265 INTERRUPTS_MASKED_CYCLES CPU1 23,911,872,654 CYCLES_NO_FPU_OPS_RETIRED
So it looks like your Opteron continues to count CYCLES_NO_FPU_OPS_RETIRED at full clock speed even when halted.
My Turion appears to slow down to ~40 MHz when halted and counts those events at that speed. Using idle=poll shows no slowdown, as expected, and unhalted clocks equals cycles_no_fpu_ops_retired.
Signature · 2 lines
Chuck "You can't read a newspaper if you can't read." --George W. Bush
Chuck,
Quoted 18 lines · Chuck Ebbert
In-Reply-To: <[email protected]>
Quoted 13 lines · Stephan Eranian
Here is what I get on my dual 2.4GHz Opteron 250:
booted with idle=halt
$ pfmon --us-c -ecpu_clk_unhalted,interrupts_masked_cycles_with_interrupt_pending,interrupts_masked_cycles,cycles_no_fpu_ops_retired -k --system-wide -t 10 <session to end in 10 seconds> CPU0 11,356,210 CPU_CLK_UNHALTED CPU0 0 INTERRUPTS_MASKED_CYCLES_WITH_INTERRUPT_PENDING CPU0 3,836,107 INTERRUPTS_MASKED_CYCLES CPU0 23,910,784,532 CYCLES_NO_FPU_OPS_RETIRED CPU1 19,303,632 CPU_CLK_UNHALTED CPU1 0 INTERRUPTS_MASKED_CYCLES_WITH_INTERRUPT_PENDING CPU1 13,942,265 INTERRUPTS_MASKED_CYCLES CPU1 23,911,872,654 CYCLES_NO_FPU_OPS_RETIRED
Quoted 3 lines
So it looks like your Opteron continues to count
CYCLES_NO_FPU_OPS_RETIREDat full clock speed even when halted.
Yes, it certainly looks like it, idling at full clock speed. But that's a server CPU! I think this may have to do with latency to get back to useful work.
Quoted 4 lines
My Turion appears to slow down to ~40 MHz when halted and counts those events at that speed. Using idle=poll shows no slowdown, as expected, and unhalted clocks equals
cycles_no_fpu_ops_retired.
That's a mobile processor, so I would expect it to use more aggressive power-saving scheme while idle, at the cost of higher latency to get back to full speed.
The good thing about this is that it was not too hard to figure out what is going using pfmon. Where you specifically looking for this? Is this a documented behavior of the Turion64?
Signature · 1 line
-Stephane