Merge pull request #763 from ryantimwilson/rusty-default-weights-fix

[rusty] Fix load stats when host is under-utilized

rust/scx_utils/src/lib.rs

@@ -83,4 +83,5 @@ pub use log_recorder::LogRecorderBuilder;
 
 mod misc;
 pub use misc::monitor_stats;
+pub use misc::normalize_load_metric;
 pub use misc::set_rlimit_infinity;

rust/scx_utils/src/misc.rs

@@ -90,3 +90,25 @@ pub fn read_file_usize(path: &Path) -> Result<usize> {
         }
     }
 }
+
+/* Load is reported as weight * duty cycle
+ *
+ * In the Linux kernel, EEDVF uses default weight = 1 s.t.
+ * load for a nice-0 thread runnable for time slice = 1
+ *
+ * To conform with cgroup weights convention, sched-ext uses
+ * the convention of default weight = 100 with the formula
+ * 100 * nice ^ 1.5. This means load for a nice-0 thread
+ * runnable for time slice = 100.
+ *
+ * To ensure we report load metrics consistently with the Linux
+ * kernel, we divide load by 100.0 prior to reporting metrics.
+ * This is also more intuitive for users since 1 CPU roughly
+ * means 1 unit of load.
+ *
+ * We only do this prior to reporting as its easier to work with
+ * weight as integers in BPF / userspace than floating point.
+ */
+pub fn normalize_load_metric(metric: f64) -> f64 {
+    metric / 100.0
+}

scheds/rust/scx_layered/src/stats.rs

@@ -18,6 +18,7 @@ use log::warn;
 use scx_stats::prelude::*;
 use scx_stats_derive::stat_doc;
 use scx_stats_derive::Stats;
+use scx_utils::normalize_load_metric;
 use serde::Deserialize;
 use serde::Serialize;
 
@@ -182,7 +183,7 @@ impl LayerStats {
         Self {
             util: stats.layer_utils[lidx] * 100.0,
             util_frac: calc_frac(stats.layer_utils[lidx], stats.total_util),
-            load: stats.layer_loads[lidx],
+            load: normalize_load_metric(stats.layer_loads[lidx]),
             load_adj: calc_frac(stats.layer_load_sums[lidx], stats.total_load_sum),
             dcycle: calc_frac(stats.layer_dcycle_sums[lidx], stats.total_dcycle_sum),
             load_frac: calc_frac(stats.layer_loads[lidx], stats.total_load),
@@ -411,7 +412,7 @@ impl SysStats {
             proc_ms: stats.processing_dur.as_millis() as u64,
             busy: stats.cpu_busy * 100.0,
             util: stats.total_util * 100.0,
-            load: stats.total_load,
+            load: normalize_load_metric(stats.total_load),
             fallback_cpu: fallback_cpu as u32,
             layers: BTreeMap::new(),
         })

scheds/rust/scx_rusty/src/load_balance.rs

@@ -155,6 +155,7 @@ use crate::stats::DomainStats;
 use crate::stats::NodeStats;
 use crate::DomainGroup;
 
+const DEFAULT_WEIGHT: f64 = bpf_intf::consts_LB_DEFAULT_WEIGHT as f64;
 const RAVG_FRAC_BITS: u32 = bpf_intf::ravg_consts_RAVG_FRAC_BITS;
 
 fn now_monotonic() -> u64 {
@@ -446,20 +447,16 @@ impl NumaNode {
     }
 
     fn stats(&self) -> NodeStats {
-        let mut stats = NodeStats {
+        let mut stats = NodeStats::new(
-            load: self.load.load_sum(),
+            self.load.load_sum(),
-            imbal: self.load.imbal(),
+            self.load.imbal(),
-            delta: self.load.delta(),
+            self.load.delta(),
-            doms: BTreeMap::new(),
+            BTreeMap::new(),
-        };
+        );
         for dom in self.domains.iter() {
             stats.doms.insert(
                 dom.id,
-                DomainStats {
+                DomainStats::new(dom.load.load_sum(), dom.load.imbal(), dom.load.delta()),
-                    load: dom.load.load_sum(),
-                    imbal: dom.load.imbal(),
-                    delta: dom.load.delta(),
-                },
             );
         }
         stats
@@ -542,8 +539,13 @@ impl<'a, 'b> LoadBalancer<'a, 'b> {
 
         let (dom_loads, total_load) = if !self.lb_apply_weight {
             (
-                ledger.dom_dcycle_sums().to_vec(),
+                ledger
-                ledger.global_dcycle_sum(),
+                    .dom_dcycle_sums()
+                    .to_vec()
+                    .into_iter()
+                    .map(|d| DEFAULT_WEIGHT * d)
+                    .collect(),
+                DEFAULT_WEIGHT * ledger.global_dcycle_sum(),
             )
         } else {
             self.infeas_threshold = ledger.effective_max_weight();
@@ -696,10 +698,12 @@ impl<'a, 'b> LoadBalancer<'a, 'b> {
                     RAVG_FRAC_BITS,
                 );
 
-                if self.lb_apply_weight {
+                let weight = if self.lb_apply_weight {
-                    let weight = (task_ctx.weight as f64).min(self.infeas_threshold);
+                    (task_ctx.weight as f64).min(self.infeas_threshold)
+                } else {
+                    DEFAULT_WEIGHT
+                };
                 load *= weight;
-                }
 
                 dom.tasks.insert(TaskInfo {
                     tptr,

scheds/rust/scx_rusty/src/stats.rs

@@ -12,6 +12,7 @@ use chrono::Local;
 use scx_stats::prelude::*;
 use scx_stats_derive::stat_doc;
 use scx_stats_derive::Stats;
+use scx_utils::normalize_load_metric;
 use scx_utils::Cpumask;
 use serde::Deserialize;
 use serde::Serialize;
@@ -39,6 +40,14 @@ pub struct DomainStats {
 }
 
 impl DomainStats {
+    pub fn new(load: f64, imbal: f64, delta: f64) -> Self {
+        Self {
+            load: normalize_load_metric(load),
+            imbal: normalize_load_metric(imbal),
+            delta: normalize_load_metric(delta),
+        }
+    }
+
     pub fn format<W: Write>(&self, w: &mut W, id: usize) -> Result<()> {
         writeln!(
             w,
@@ -67,6 +76,15 @@ pub struct NodeStats {
 }
 
 impl NodeStats {
+    pub fn new(load: f64, imbal: f64, delta: f64, doms: BTreeMap<usize, DomainStats>) -> Self {
+        Self {
+            load: normalize_load_metric(load),
+            imbal: normalize_load_metric(imbal),
+            delta: normalize_load_metric(delta),
+            doms,
+        }
+    }
+
     pub fn format<W: Write>(&self, w: &mut W, id: usize) -> Result<()> {
         writeln!(
             w,