LeafNodeRank: kOutboundFanoutWindow=16, driver_threads=nproc/4 on SMT-on (#734)

Summary:
Pull Request resolved: #734

Two small platform-tuning changes:

1. LeafNodeRank.cc: kOutboundFanoutWindow 32 → 16. With nproc-sized
   ThriftSrv.IO pool (D105119223) and per-MockClient EventBase
   placement, 32 was overbatching outbound RPCs per request and
   inflating dispatch_per_rpc tail latency on BGM/Turin (88-176
   cores). 16 keeps a sane queue depth on big-SMT hosts without
   meaningful throughput loss on smaller boxes.

2. run.sh: DRIVER_THREADS unified at nproc/4 across SMT-on and SMT-off.
   Previously SMT-on used nproc/5, SMT-off used nproc/4 — an
   inconsistency that under-pinned driver work on big SMT hosts (e.g.
   BGM 176 logical: 35 SMT-on driver threads vs 44 SMT-off). Using
   nproc/4 in both branches removes one source of cross-platform
   variance.

Reviewed By: YifanYuan3

Differential Revision: D105659809

fbshipit-source-id: e276e98b443f2aa447c4bd53edd37f717abce92a

Author: excelle08

packages/feedsim/run.sh

@@ -50,15 +50,16 @@ EVENTBASE_THREADS_DEFAULT="$(nproc)"  # nproc threads so the ThriftSrv.IO pool i
 SRV_THREADS_DEFAULT=8        # 8 should also suffice for most purposes
 if [[ "$IS_SMT_ON" = 1 ]]; then
   RANKING_THREADS_DEFAULT="$(( $(nproc) * 7/20))"  # 7/20 is 0.35 cpu factor
-  DRIVER_THREADS="$(echo "scale=2; $(nproc) / 5.0 + 0.5 " | bc )"  # Driver threads, rounds nearest.
-  DRIVER_THREADS="${DRIVER_THREADS%.*}"  # Truncate decimal fraction.
-  DRIVER_THREADS="$(echo "${BC_MAX_FN}; max(${DRIVER_THREADS:-0}, 4)" | bc )" # At least 4 threads.
 else
   RANKING_THREADS_DEFAULT="$(( $(nproc) * 15/20))"  # 15/20 is 0.75 cpu factor
-  DRIVER_THREADS="$(echo "scale=2; $(nproc) / 4.0 + 0.5 " | bc )"  # Driver threads, rounds nearest.
-  DRIVER_THREADS="${DRIVER_THREADS%.*}"  # Truncate decimal fraction.
-  DRIVER_THREADS="$(echo "${BC_MAX_FN}; max(${DRIVER_THREADS:-0}, 4)" | bc )" # At least 4 threads.
 fi
+# Driver threads = nproc/4 across SMT-on / SMT-off. Previously the SMT-on
+# branch used nproc/5, which under-pinned driver work on big SMT boxes
+# (e.g. BGM 176 logical → 35 vs 44 threads). Standardizing both branches
+# at nproc/4 removes one source of cross-platform variance.
+DRIVER_THREADS="$(echo "scale=2; $(nproc) / 4.0 + 0.5 " | bc )"  # rounds nearest
+DRIVER_THREADS="${DRIVER_THREADS%.*}"  # Truncate decimal fraction.
+DRIVER_THREADS="$(echo "${BC_MAX_FN}; max(${DRIVER_THREADS:-0}, 4)" | bc )" # At least 4 threads.
 
 show_help() {
 cat <<EOF

packages/feedsim/third_party/src/workloads/ranking/LeafNodeRank.cc

@@ -728,11 +728,7 @@ static feedsim::LatencyHistogram g_sampled_lat_us;
 // dispatch_per_rpc averages ~500ms even though mock_handler_actual was
 // ~4ms — pure EB queue depth. window(K) issues K up-front and then
 // refills as each completes, capping per-request EB pressure.
-//
-// 32 was picked to be small enough to keep the EB queue shallow (each
-// in-flight RPC ~5-200ms means the queue completes in 50-300ms) while
-// still amortizing the per-RPC overhead of the window machinery.
-static constexpr size_t kOutboundFanoutWindow = 32;
+static constexpr size_t kOutboundFanoutWindow = 16;
 
 namespace {
 struct FanoutSpec {