Profile steady-state block workloads

Author: Arbousier1

.github/workflows/phase2-runtime-ab.yml

@@ -38,11 +38,11 @@ on:
         default: "180"
         type: string
       spark_profile_mode:
-        description: "Optional 4-run Spark diagnostic for block-direct-write"
+        description: "Optional 4-run Spark diagnostic (cpu=slow ticks, cpu-all=steady-state)"
         required: true
         default: "none"
         type: choice
-        options: ["none", "cpu", "alloc"]
+        options: ["none", "cpu", "cpu-all", "alloc"]
 
 permissions:
   contents: read
@@ -139,9 +139,15 @@ jobs:
           [[ "$WARMUP_SECONDS" =~ ^[0-9]+$ ]] && (( WARMUP_SECONDS >= 10 ))
           [[ "$SETTLE_SECONDS" =~ ^[0-9]+$ ]] && (( SETTLE_SECONDS >= 5 ))
           [[ "$MEASURE_SECONDS" =~ ^[0-9]+$ ]] && (( MEASURE_SECONDS >= 10 ))
-          [[ "$SPARK_PROFILE_MODE" =~ ^(none|cpu|alloc)$ ]]
-          if [[ "$SPARK_PROFILE_MODE" != none && "$SCENARIO" != block-direct-write ]]; then
-            echo "Spark profiling is currently isolated to block-direct-write" >&2
+          [[ "$SPARK_PROFILE_MODE" =~ ^(none|cpu|cpu-all|alloc)$ ]]
+          if [[ "$SPARK_PROFILE_MODE" == cpu-all && \
+                "$SCENARIO" != block-active && "$SCENARIO" != block-direct-write ]]; then
+            echo "Spark cpu-all profiling is isolated to block-active or block-direct-write" >&2
+            exit 64
+          fi
+          if [[ "$SPARK_PROFILE_MODE" != none && "$SPARK_PROFILE_MODE" != cpu-all && \
+                "$SCENARIO" != block-direct-write ]]; then
+            echo "Spark cpu/alloc profiling is isolated to block-direct-write" >&2
             exit 64
           fi
           if [[ "$SPARK_PROFILE_MODE" != none && "$RUNS" != 4 ]]; then

tools/perf/run-phase2-runtime-once.sh

@@ -86,11 +86,17 @@ fi
 [[ "$protocol_trace_enabled" == 0 || "$protocol_trace_enabled" == 1 ]] \
   || { echo "PHASE2_PROTOCOL_TRACE_ENABLED must be 0 or 1" >&2; exit 64; }
 case "$spark_profile_mode" in
-  none|cpu|alloc) ;;
-  *) echo "PHASE2_SPARK_PROFILE_MODE must be none, cpu, or alloc" >&2; exit 64 ;;
+  none|cpu|cpu-all|alloc) ;;
+  *) echo "PHASE2_SPARK_PROFILE_MODE must be none, cpu, cpu-all, or alloc" >&2; exit 64 ;;
 esac
-if [[ "$spark_profile_mode" != none && "$scenario" != block-direct-write ]]; then
-  echo "Spark profiling is currently isolated to block-direct-write" >&2
+if [[ "$spark_profile_mode" == cpu-all && \
+      "$scenario" != block-active && "$scenario" != block-direct-write ]]; then
+  echo "Spark cpu-all profiling is isolated to block-active or block-direct-write" >&2
+  exit 64
+fi
+if [[ "$spark_profile_mode" != none && "$spark_profile_mode" != cpu-all && \
+      "$scenario" != block-direct-write ]]; then
+  echo "Spark cpu/alloc profiling is isolated to block-direct-write" >&2
   exit 64
 fi
 [[ -f "$plugin_jar" && -f "$paper_jar" ]] \
@@ -638,6 +644,12 @@ if [[ "$spark_profile_mode" != none ]]; then
       spark_profile_interval_unit="milliseconds"
       spark_profile_only_ticks_over_ms="40"
       ;;
+    cpu-all)
+      spark_profile_start_command="spark profiler start --interval 1"
+      spark_profile_interval="1"
+      spark_profile_interval_unit="milliseconds"
+      spark_profile_only_ticks_over_ms=""
+      ;;
     alloc)
       spark_profile_start_command="spark profiler start --alloc --interval 32768"
       spark_profile_interval="32768"