ci(bench): add IRONCACHE_SHARDS probe knob to the head-to-head

Add an optional `ironcache_shards` workflow input (and the matching IRONCACHE_SHARDS
env override in headtohead.sh) that DECOUPLES IronCache's shard (= runtime-thread)
count from the pinned core count. Blank keeps the thread-per-core norm (one shard per
pinned core). This isolates thread-oversubscription effects: e.g. run 1 shard vs 2
shards on the SAME 2 pinned cores and compare raw qps. The qps_per_core denominator
stays the pinned core count, so a probe reads RAW qps across shard counts. Diagnostic
only; no change to the default behavior.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: Zeke <ezequiel.lares@outlook.com>

Author: ELares

.github/workflows/headtohead.yml

@@ -39,6 +39,10 @@ on:
         description: "Measured closed-loop pass duration (seconds)."
         type: string
         default: "8"
+      ironcache_shards:
+        description: "Override IronCache shard (= runtime-thread) count; blank = one per pinned core. Probes thread-oversubscription (e.g. 1 shard on 2 pinned cores)."
+        type: string
+        default: ""
 
 permissions:
   contents: read
@@ -100,6 +104,7 @@ jobs:
           KEYSPACE: ${{ inputs.keycount }}
           DURATION_SECS: ${{ inputs.duration_secs }}
           WARMUP_SECS: "3"
+          IRONCACHE_SHARDS: ${{ inputs.ironcache_shards }}
         run: |
           set -euo pipefail
           # The harness resolves the competitor (valkey-server, else redis-server),

scripts/bench/headtohead.sh

@@ -474,10 +474,16 @@ measure_server() {
 
   # (b) Boot it, pinned, persistence OFF + eviction OFF for the measurement.
   if [[ "${kind}" == "ironcache" ]]; then
-    echo "[h2h] starting ${name} on ${HOST}:${PORT} (shards=${SERVER_CORE_COUNT}, maxmemory=${MAXMEMORY})..."
+    # IRONCACHE_SHARDS lets a probe DECOUPLE the shard (= runtime-thread) count from the
+    # pinned core count, to isolate thread-oversubscription effects (e.g. 1 shard vs 2
+    # shards on the SAME 2 pinned cores). Defaults to one shard per pinned core (the
+    # thread-per-core norm). The qps_per_core denominator stays SERVER_CORE_COUNT (cores),
+    # so a probe compares RAW qps across shard counts.
+    local ic_shards="${IRONCACHE_SHARDS:-${SERVER_CORE_COUNT}}"
+    echo "[h2h] starting ${name} on ${HOST}:${PORT} (shards=${ic_shards} over ${SERVER_CORE_COUNT} pinned cores, maxmemory=${MAXMEMORY})..."
     IRONCACHE_MAXMEMORY="${MAXMEMORY}" \
       ${SERVER_PREFIX[@]+"${SERVER_PREFIX[@]}"} "${IRONCACHE_BIN}" \
-      --port "${PORT}" --shards "${SERVER_CORE_COUNT}" server \
+      --port "${PORT}" --shards "${ic_shards}" server \
       >"${SERVER_LOG}" 2>&1 &
     SERVER_PID=$!
   else