On 24-Jul-19 2:18 PM, David Hunt wrote:
The branch ratio algorithm in the vm_power_manager sample application
can be very sensitive at patricular loads in a workload, causing
oscillations between min and max frequency. For example, if a
workload is at 50%, scaling up may change the ratio
enough that it immediately thinks it needs to scale down again.
This patch introduces a sliding window recording the scale up/down
direction for the last 32 samples, and scales up if any samples indicate
we should scale up, otherwise scale down. Each core has it's own window.
Fixes: 4b1a631b8a8a ("examples/vm_power: add oob monitoring functions")
Cc: sta...@dpdk.org
Signed-off-by: David Hunt <david.h...@intel.com>
---
examples/vm_power_manager/oob_monitor_x86.c | 34 +++++++++++++++++++--
examples/vm_power_manager/power_manager.c | 3 +-
examples/vm_power_manager/power_manager.h | 12 ++++++++
3 files changed, 44 insertions(+), 5 deletions(-)
diff --git a/examples/vm_power_manager/oob_monitor_x86.c
b/examples/vm_power_manager/oob_monitor_x86.c
index ebd96b205..aecfcb2eb 100644
--- a/examples/vm_power_manager/oob_monitor_x86.c
+++ b/examples/vm_power_manager/oob_monitor_x86.c
@@ -39,6 +39,7 @@ apply_policy(int core)
int64_t hits_diff, miss_diff;
float ratio;
int ret;
+ int freq_window_idx, up_count = 0, i;
g_active = 0;
ci = get_core_info();
@@ -101,10 +102,37 @@ apply_policy(int core)
ratio = (float)miss_diff * (float)100 / (float)hits_diff;
- if (ratio < ci->branch_ratio_threshold)
- power_manager_scale_core_min(core);
+ /*
+ * Store the last few directions that the ratio indicates
+ * we should take. If there's on 'up', then we scale up
One?
+ * quickly. If all indicate 'down', only then do we scale
+ * down. Each core_details struct has it's own array.
+ */
+ freq_window_idx = ci->cd[core].freq_window_idx;
+ if (ratio > ci->branch_ratio_threshold)
+ ci->cd[core].freq_directions[freq_window_idx] = 1;
else
- power_manager_scale_core_max(core);
+ ci->cd[core].freq_directions[freq_window_idx] = 0;
+
+ freq_window_idx++;
+ freq_window_idx = freq_window_idx & (FREQ_WINDOW_SIZE-1);
+ ci->cd[core].freq_window_idx = freq_window_idx;
+
+ up_count = 0;
+ for (i = 0; i < FREQ_WINDOW_SIZE; i++)
+ up_count += ci->cd[core].freq_directions[i];
+
+ if (up_count == 0) {
+ if (ci->cd[core].freq_state != FREQ_MIN) {
+ power_manager_scale_core_min(core);
+ ci->cd[core].freq_state = FREQ_MIN;
+ }
+ } else {
+ if (ci->cd[core].freq_state != FREQ_MAX) {
+ power_manager_scale_core_max(core);
+ ci->cd[core].freq_state = FREQ_MAX;
+ }
+ }
So it's biased towards scaling up quickly, but it's doing that over a
period. Please correct me if i'm wrong as i'm not really familiar with
this codebase, but, assuming the window size is long enough, you could
be missing opportunities to scale down? For example, if you get a short
burst of 1's followed by a long burst of zeroes, you're not scaling down
until you go through the entire buffer and overwrite all of the values.
I guess that's the point of oscillation prevention, but maybe you could
improve the "scale-up" part by only checking a few recent values, rather
than the entire buffer?
--
Thanks,
Anatoly
