swaits/hsmc
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs(bench): refresh baseline.txt with post-opt numbers and rationale

Browse Source 
Captures Opt #1-#4 outcomes:
- Opt #2 (get_unchecked) and #4 (timer-pop inline) landed.
- Opt #3 (run_action proxy/context) was already free.
- Opt #1 (Option<u16> → sentinel) tried, reverted (-0.6% net).

Documents the gungraun-wrapper measurement floor.
This commit is contained in:
swaits committed 2026-05-04 15:29:24 -06:00
1 parent 8cd1aea646
commit 2f37f261c5
1 file changed
+149 -90
Download Patch File Download Diff File Expand all files Collapse all files
hsmc/benches/baseline.txt
+149 -90
View File
@@ -2,119 +2,163 @@
# #
# ─── Harness migration (iai-callgrind 0.16 → gungraun 0.18) ────────── # ─── Harness migration (iai-callgrind 0.16 → gungraun 0.18) ──────────
# #
# Switched from iai-callgrind to gungraun (a maintained fork). The # Switched from iai-callgrind to gungraun (the renamed continuation —
# gungraun harness wraps each bench body with a slightly thicker shim # crates.io shows iai-callgrind frozen at 0.16.1; the gh repo
# (extra `__gungraun_wrapper_id_mod_<id>::<id>` indirection, more # iai-callgrind/iai-callgrind redirects to gungraun/gungraun). Same
# careful black_box / setup-result handoff), so absolute instruction # valgrind backend, same callgrind output format. Old named baselines
# counts are NOT comparable to the iai-callgrind 0.16 numbers in git # (pre_lca_fix, before_*) are gone; their deltas live in the historical
# history. The HSM dispatch work itself didn't change between the two # block at the bottom for context.
# runs — the harness is just instrumenting more.
# #
# Below are the gungraun_initial baselines. Old named baselines # Re-run: cargo bench --bench dispatch
# (pre_lca_fix, before_path_precompute, …) are gone; their deltas live # Diff: cargo bench --bench dispatch -- --baseline=gungraun_initial
# in the historical comment block at the bottom for context, but they # Snapshot: cargo bench --bench dispatch -- --save-baseline=before_<change>
# are not directly comparable to current numbers.
#
# Re-run with:
#
# cargo bench --bench dispatch
#
# Save a named baseline for diffing:
#
# cargo bench --bench dispatch -- --save-baseline=before_<your_change>
# <change>
# cargo bench --bench dispatch -- --baseline=before_<your_change>
# #
# Numbers below are from a 6.19.14-zen kernel x86_64 box, valgrind # Numbers below are from a 6.19.14-zen kernel x86_64 box, valgrind
# 3.25.1, gungraun 0.18.2, gungraun-runner 0.18.2, --release. # 3.25.1, gungraun 0.18.2, gungraun-runner 0.18.2, --release.
#
# Saved baselines (all under target/gungraun/...):
# gungraun_initial first gungraun-harness run, post-migration
# after_opt2_unchecked get_unchecked for codegen table indexing
# after_opt4_timer_inline pop_expired_nonempty inline (current numbers)
# before_more_opts alias for after_opt4_timer_inline (sweep base)
dispatch::dispatch::h2_cold_start cold:setup_h2_cold() dispatch::dispatch::h2_cold_start cold:setup_h2_cold()
Instructions: 131 Instructions: 24
L1 Hits: 181 L1 Hits: 39
RAM Hits: 12 RAM Hits: 3
Estimated Cycles: 601 Estimated Cycles: 144
dispatch::dispatch::h2_warm_lateral warm:setup_h2_warm() dispatch::dispatch::h2_warm_lateral warm:setup_h2_warm()
Instructions: 119 Instructions: 53
L1 Hits: 168 L1 Hits: 77
RAM Hits: 7 RAM Hits: 4
Estimated Cycles: 413 Estimated Cycles: 217
dispatch::dispatch::h4_cross_tree cross:setup_h4_warm() dispatch::dispatch::h4_cross_tree cross:setup_h4_warm()
Instructions: 128 Instructions: 93
L1 Hits: 182 L1 Hits: 130
RAM Hits: 7 RAM Hits: 10
Estimated Cycles: 427 Estimated Cycles: 480
dispatch::dispatch::h4_up up:setup_h4_warm() dispatch::dispatch::h4_up up:setup_h4_warm()
Instructions: 114 Instructions: 88
L1 Hits: 163 L1 Hits: 125
RAM Hits: 6 RAM Hits: 8
Estimated Cycles: 373 Estimated Cycles: 405
dispatch::dispatch::h4_self self_:setup_h4_warm() dispatch::dispatch::h4_self self_:setup_h4_warm()
Instructions: 128 Instructions: 93
L1 Hits: 182 L1 Hits: 130
RAM Hits: 7 RAM Hits: 10
Estimated Cycles: 427 Estimated Cycles: 480
dispatch::dispatch::h8_cross_tree cross:setup_h8_warm() dispatch::dispatch::h8_cross_tree cross:setup_h8_warm()
Instructions: 97 Instructions: 53
L1 Hits: 137 L1 Hits: 76
RAM Hits: 8 RAM Hits: 5
Estimated Cycles: 417 Estimated Cycles: 251
dispatch::dispatch::bubble_leaf leaf:setup_bubble_warm() dispatch::dispatch::bubble_leaf leaf:setup_bubble_warm()
Instructions: 129 Instructions: 94
L1 Hits: 183 L1 Hits: 132
RAM Hits: 9 RAM Hits: 11
Estimated Cycles: 498 Estimated Cycles: 517
dispatch::dispatch::bubble_mid mid:setup_bubble_warm() dispatch::dispatch::bubble_mid mid:setup_bubble_warm()
Instructions: 129 Instructions: 94
L1 Hits: 183 L1 Hits: 133
RAM Hits: 9 RAM Hits: 10
Estimated Cycles: 498 Estimated Cycles: 483
dispatch::dispatch::bubble_top top:setup_bubble_warm() dispatch::dispatch::bubble_top top:setup_bubble_warm()
Instructions: 129 Instructions: 93
L1 Hits: 184 L1 Hits: 132
RAM Hits: 8 RAM Hits: 10
Estimated Cycles: 464
dispatch::dispatch::timer_fire_dispatch fire:setup_timer_fire_armed()
Instructions: 221
L1 Hits: 294
RAM Hits: 12
Estimated Cycles: 714
dispatch::dispatch::terminate_event quit:setup_term_warm()
Instructions: 68
L1 Hits: 99
RAM Hits: 7
Estimated Cycles: 344
dispatch::dispatch::bubble_cross_tree switch:setup_bubble_cross_warm()
Instructions: 120
L1 Hits: 167
RAM Hits: 9
Estimated Cycles: 482 Estimated Cycles: 482
dispatch::dispatch::chain_self_emit chain:setup_chain_warm() dispatch::dispatch::timer_fire_dispatch fire:setup_timer_fire_armed()
Instructions: 500 Instructions: 175
L1 Hits: 689 L1 Hits: 237
RAM Hits: 18 RAM Hits: 7
Estimated Cycles: 1319 Estimated Cycles: 482
dispatch::dispatch::terminate_event quit:setup_term_warm()
Instructions: 53
L1 Hits: 81
RAM Hits: 5
Estimated Cycles: 256
dispatch::dispatch::bubble_cross_tree switch:setup_bubble_cross_warm()
Instructions: 85
L1 Hits: 116
RAM Hits: 11
Estimated Cycles: 501
dispatch::dispatch::chain_self_emit chain:setup_chain_warm()
Instructions: 439
L1 Hits: 613
RAM Hits: 17
Estimated Cycles: 1213
# ─── Migration deltas (gungraun_initial → current) ──────────────────
#
# Two optimization passes landed under gungraun:
#
# Opt #2: get_unchecked for codegen table indexing
# __DISPATCH, __DEFAULT_CHILD, __PATH_RANGE/_DATA, __TERMINATE_*
# all index by codegen-emitted state ids that are bounded by
# construction. Replacing the [idx] / [s..e] checked forms with
# get_unchecked unlocked aggressive inlining beyond what the
# bounds-check elision alone would buy: apply_path inlines into
# step() for h2/h4/h8/bubble; step() inlines into the bench
# wrapper for h2_cold_start / h2_warm_lateral / h8_cross_tree.
# The cold-path enter chain on h2_cold_start collapses to one
# u16 store. Bench impact:
# h2_cold_start 131 → 24 -82%
# h2_warm_lateral 119 → 53 -55%
# h8_cross_tree 97 → 53 -45%
# bubble_cross_tree 120 → 85 -29%
# h4_cross_tree/_self 128 → 93 -27%
# bubble_leaf/mid/top 129 → 94 -27%
# h4_up 114 → 88 -23%
# terminate_event 68 → 53 -22%
# chain_self_emit 500 → 439 -12%
# timer_fire_dispatch 221 → 212 -4%
#
# Opt #4: TimerTable::pop_expired_nonempty inline
# Drop #[cold], add #[inline]. The cold hint was forcing the body
# out-of-line at every step()-with-timers call site, paying full
# x86_64 ABI prologue/epilogue (~10-15 instr) per call. Inlining
# also let LLVM fold run_handlers into step() in the timer path
# (knock-on effect: the inliner wasn't budget-blocked by a separate
# pop_expired_nonempty call). Bench impact:
# timer_fire_dispatch 212 → 175 -17%
# (others unchanged)
#
# Tried but reverted:
#
# Opt #1: current: Option<u16> → u16 with u16::MAX sentinel.
# Theory: drop the discriminant byte, single-load is_none-check
# vs tag-then-value, one-write state-set. In practice net -9 instr
# across all 13 benches (-0.6%): real wins on terminate_event (-6)
# and bubble_cross_tree (-10) where the writes dominate, offset
# by compiler-inliner noise (+2 to +3 instr) on h4_cross_tree,
# h4_self, bubble_mid, bubble_top, chain_self_emit. Aggregate gain
# not worth the maintenance burden of sentinel-aware code at every
# self.current site plus an Option<u16> conversion helper for the
# observation API.
#
# Opt #3: hand-tighten run_action's EmitProxy/ActionContext init.
# Inspection of the compiled bench binary confirmed run_action,
# EmitProxy, ActionContext, and even the chain bench's emit_next
# vtable dispatch are all already fully inlined — no separate
# symbol exists for any of them. Already free; no change made.
#
# ─── Historical context (iai-callgrind era, NOT comparable) ───────── # ─── Historical context (iai-callgrind era, NOT comparable) ─────────
# #
# The codebase went through six dispatch-path optimization passes # The codebase went through six dispatch-path optimization passes
# under iai-callgrind 0.16 before this harness migration. Cumulative # under iai-callgrind 0.16 before the harness migration. Cumulative
# results vs. the original O(H²) LCA scan: # h8_cross_tree result vs. the original O(H²) LCA scan: -82% instr.
#
# h8_cross_tree -82% instr / -72% cycles (deepest LCA stress)
# terminate_event -75% instr (combined) (precomputed exit table)
#
# The passes, in order: # The passes, in order:
# #
# 1. LCA contains() → O(H) ancestor-walk # 1. LCA contains() → O(H) ancestor-walk
@@ -125,7 +169,8 @@ dispatch::dispatch::chain_self_emit chain:setup_chain_warm()
# 6. Precomputed __TERMINATE_DATA + queue.clear() + u128→u64 nanoseconds # 6. Precomputed __TERMINATE_DATA + queue.clear() + u128→u64 nanoseconds
# #
# Pre-migration absolute numbers (iai-callgrind 0.16, for reference # Pre-migration absolute numbers (iai-callgrind 0.16, for reference
# only; do not subtract from gungraun numbers): # only; not directly comparable to gungraun numbers because the
# gungraun bench wrapper instruments slightly more):
# #
# h2_cold_start 38 instr 201 cycles # h2_cold_start 38 instr 201 cycles
# h2_warm_lateral 84 instr 397 cycles # h2_warm_lateral 84 instr 397 cycles
@@ -139,7 +184,21 @@ dispatch::dispatch::chain_self_emit chain:setup_chain_warm()
# bubble_cross_tree 94 instr 444 cycles # bubble_cross_tree 94 instr 444 cycles
# chain_self_emit 419 instr 1301 cycles # chain_self_emit 419 instr 1301 cycles
# #
# Most of the absolute-instruction increase under gungraun is # Bench coverage:
# harness, not HSM. Comparable iai-callgrind/gungraun deltas (e.g., #
# h8_cross_tree 93→97, terminate_event 69→68) confirm the dispatch # `timer_fire_dispatch` — only bench that exercises pop_expired's
# work itself is unchanged. # non-empty branch and `dispatch_trigger`.
# `terminate_event` — only bench that exercises `__check_terminate`'s
# true arm and `do_terminate`.
# `bubble_cross_tree` — handler on L1 (4 bubble hops) firing a
# 10-state cross-tree transition. Realistic HSM worst case.
#
# ─── Measurement floor ───────────────────────────────────────────────
#
# After Opt #2, the gungraun bench-harness wrapper is the dominant
# fixed cost on simple benches (~28-49 instr per bench). On
# h2_cold_start and h8_cross_tree, the entire bench body is inlined
# into the wrapper — at 24 and 53 instructions respectively, you are
# looking at the harness floor plus 0-25 instr of HSM work. Future
# delta tables should keep this in mind: a 5-instr "win" on h2 is
# 20% of the HSM-attributable cost, not 5/53 = 9%.