RE: [tip:x86/urgent] x86 idle: Repair large-server 50-watt idle-power regression

[Brown, Len]

Davidlohr,

Thanks for the note.

Ideally (on Linux in general, and on servers, in particular) we strive
for the performance impact of power saving features to be small enough
to be considered in "measurement noise".

Your report for 160 core Westmere AIM numbers being hit at 10-25%
shows 15% measurement noise?  But even if true, this looks bad.

Any chance you can re-run, with the following two tweaks,
one at a time?

I'd be curious if you can wrap the invocation in turbostat -v
and capture that output to how what states we are seeing
during the benchmark run.

thanks,
-Len



#1: skip flush for C1
        
diff --git a/drivers/idle/intel_idle.c b/drivers/idle/intel_idle.c
index f80b700..6027d06 100644
--- a/drivers/idle/intel_idle.c
+++ b/drivers/idle/intel_idle.c
@@ -377,7 +377,7 @@ static int intel_idle(struct cpuidle_device *dev,
 
        if (!current_set_polling_and_test()) {
 
-               if (this_cpu_has(X86_FEATURE_CLFLUSH_MONITOR))
+               if ((eax > 0) && this_cpu_has(X86_FEATURE_CLFLUSH_MONITOR))
                        clflush((void *)&current_thread_info()->flags);
 
                __monitor((void *)&current_thread_info()->flags, 0, 0);


#2: skip flush for C1 and C1E
        
diff --git a/drivers/idle/intel_idle.c b/drivers/idle/intel_idle.c
index f80b700..6027d06 100644
--- a/drivers/idle/intel_idle.c
+++ b/drivers/idle/intel_idle.c
@@ -377,7 +377,7 @@ static int intel_idle(struct cpuidle_device *dev,
 
        if (!current_set_polling_and_test()) {
 
-               if (this_cpu_has(X86_FEATURE_CLFLUSH_MONITOR))
+               if ((eax > 1) && this_cpu_has(X86_FEATURE_CLFLUSH_MONITOR))
                        clflush((void *)&current_thread_info()->flags);
 
                __monitor((void *)&current_thread_info()->flags, 0, 0);