Brummel
ac70011352
The latency_harness.py harness spawns the bench binary on a PTY, records monotonic_ns() per stdout line, and reports inter-arrival gap distribution (median / p99 / p99.9 / max). Tail latency is Decision 10's central real-time claim; total wall-time and RSS are the wrong metrics for that question. Paired fixtures: bench_latency_implicit (Boehm-fair, no mode annotations, leaks under --alloc=rc) and bench_latency_explicit (mode-annotated hot path, what RC was built for). Both use a depth-19 balanced tree (~16 MB) as the persistent live working set, plus per-op IntList build+sum churn forcing GC pressure under Boehm. Authored by ailang-bencher; ships evidence, not features.
226 lines
7.7 KiB
Plaintext
226 lines
7.7 KiB
Plaintext
; Latency-distribution bench fixture — Implicit-mode variant.
|
|
;
|
|
; Companion to bench_latency_explicit. Together they test the
|
|
; hypothesis "Boehm has unbounded p99 per-operation latency under
|
|
; continuous alloc pressure with a large persistent live working
|
|
; set; RC under explicit-mode has p99 within a small constant
|
|
; factor of the median".
|
|
;
|
|
; Implicit-mode variant: no `(borrow T)`, `(own T)`, `(reuse-as)`,
|
|
; `(drop-iterative)` annotations. This is the canonical "Boehm-fair"
|
|
; arm — the way you'd write the program without thinking about
|
|
; modes. Under `--alloc=gc`, Boehm cleans up. Under `--alloc=rc`,
|
|
; this variant LEAKS (Implicit params are not dec'd) and is not a
|
|
; meaningful RC measurement; the bench harness intentionally only
|
|
; runs this fixture under `--alloc=gc`.
|
|
;
|
|
; Workload:
|
|
; - Live cache: balanced binary tree of depth 19 (524_287 nodes,
|
|
; ~16 MB). Stays referenced through the entire bench loop.
|
|
; - Per-op work: build a 500-cell IntList of 0..499, sum it
|
|
; (sum = 124750), print one stdout marker line every PRINT_K
|
|
; ops. Total churn: 20000 * 500 cells = 10M cell-allocs ≈
|
|
; 240 MB ≫ live-set, forcing Boehm to collect many times.
|
|
; - Total ops: 20_000. Print every PRINT_K=20 ops → 1000 timing
|
|
; samples + 1 final summary line.
|
|
;
|
|
; What the harness sees:
|
|
; - One "READY" line at startup once the tree is built.
|
|
; - 1000 lines, each containing the per-chunk sum (always 124750)
|
|
; so output stays validatable. The harness ignores values and
|
|
; records only inter-arrival times.
|
|
; - One final "DONE" line.
|
|
;
|
|
; The harness times each line's arrival via clock_gettime on its
|
|
; end of a PTY-controlled stdout (PTY forces line-buffering through
|
|
; libc's printf), then computes median / p99 / p99.9 / max of the
|
|
; gaps.
|
|
;
|
|
; Why a print-driven gap measurement: AILang has no high-resolution
|
|
; clock extern. Adding one would mean a codegen change (a new `do
|
|
; bench/clock` op routed into the codegen seam), which is
|
|
; implementer territory, not bencher territory. Stdout-gap timing
|
|
; has a noise floor of ~10-50 µs (printf + pipe roundtrip) which is
|
|
; well below the millisecond-scale STW pauses the hypothesis
|
|
; predicts; if the hypothesis is right, the signal swamps the
|
|
; noise. If the data shows a tighter distribution than that noise
|
|
; floor, we'll have to escalate to in-process clocks; otherwise the
|
|
; bench is sufficient.
|
|
|
|
(module bench_latency_implicit
|
|
|
|
(data Tree
|
|
(doc "Balanced binary tree, 32-byte cells (tag + Int payload + 2 ptrs).")
|
|
(ctor TLeaf)
|
|
(ctor TNode (con Int) (con Tree) (con Tree)))
|
|
|
|
(data IntList
|
|
(doc "Singly-linked Int list, 24-byte cells.")
|
|
(ctor LNil)
|
|
(ctor LCons (con Int) (con IntList)))
|
|
|
|
; ---------- Live cache: balanced tree of given depth ----------
|
|
|
|
(fn build_tree
|
|
(doc "Build a balanced tree of given depth, every value = 1. Constructor-blocked — recursion depth = `depth`, fits 8MB stack at depth 19.")
|
|
(type
|
|
(fn-type
|
|
(params (con Int))
|
|
(ret (con Tree))))
|
|
(params depth)
|
|
(body
|
|
(if (app == depth 0)
|
|
(term-ctor Tree TLeaf)
|
|
(term-ctor Tree TNode
|
|
1
|
|
(app build_tree (app - depth 1))
|
|
(app build_tree (app - depth 1))))))
|
|
|
|
(fn sum_tree
|
|
(doc "Touch every node of the tree (ensures liveness across the loop).")
|
|
(type
|
|
(fn-type
|
|
(params (con Tree))
|
|
(ret (con Int))))
|
|
(params t)
|
|
(body
|
|
(match t
|
|
(case (pat-ctor TLeaf) 0)
|
|
(case (pat-ctor TNode v l r)
|
|
(app + v (app + (app sum_tree l) (app sum_tree r)))))))
|
|
|
|
; ---------- Per-op work: build/sum an N-cell list ----------
|
|
|
|
(fn cons_n_acc
|
|
(doc "Tail-recursive list builder. Result = [n-1, n-2, ..., 0] :: IntList.")
|
|
(type
|
|
(fn-type
|
|
(params (con Int) (con IntList))
|
|
(ret (con IntList))))
|
|
(params n acc)
|
|
(body
|
|
(if (app == n 0)
|
|
acc
|
|
(tail-app cons_n_acc
|
|
(app - n 1)
|
|
(term-ctor IntList LCons (app - n 1) acc)))))
|
|
|
|
(fn cons_n
|
|
(doc "Build [0,1,...,n-1] :: IntList.")
|
|
(type
|
|
(fn-type
|
|
(params (con Int))
|
|
(ret (con IntList))))
|
|
(params n)
|
|
(body
|
|
(app cons_n_acc n (term-ctor IntList LNil))))
|
|
|
|
(fn sum_list_acc
|
|
(doc "Tail-recursive sum.")
|
|
(type
|
|
(fn-type
|
|
(params (con IntList) (con Int))
|
|
(ret (con Int))))
|
|
(params xs acc)
|
|
(body
|
|
(match xs
|
|
(case (pat-ctor LNil) acc)
|
|
(case (pat-ctor LCons h t)
|
|
(tail-app sum_list_acc t (app + acc h))))))
|
|
|
|
(fn sum_list
|
|
(doc "Sum every element. Calls sum_list_acc with seed 0.")
|
|
(type
|
|
(fn-type
|
|
(params (con IntList))
|
|
(ret (con Int))))
|
|
(params xs)
|
|
(body
|
|
(app sum_list_acc xs 0)))
|
|
|
|
; One operation: build and sum a list of length CHUNK_LEN, return
|
|
; the sum. The tree `t` is passed through and subjected to
|
|
; `sum_tree` so the optimizer can't eliminate it, but the result
|
|
; is XOR'd back into the int we return so the value chain stays
|
|
; live without unbounded accumulation.
|
|
;
|
|
; Note: we don't actually want sum_tree to fire on every op (it
|
|
; would dominate the per-op cost and bury allocator effects).
|
|
; Instead we touch only the tree's root via a cheap `pin_root`
|
|
; that pattern-matches once. Boehm's tracing still walks the
|
|
; whole tree on every collection because the tree pointer is
|
|
; live through the loop scope.
|
|
(fn pin_root
|
|
(doc "Constant-time tree liveness pin — read root tag, return 1 (TNode) or 0 (TLeaf).")
|
|
(type
|
|
(fn-type
|
|
(params (con Tree))
|
|
(ret (con Int))))
|
|
(params t)
|
|
(body
|
|
(match t
|
|
(case (pat-ctor TLeaf) 0)
|
|
(case (pat-ctor TNode v l r) 1))))
|
|
|
|
(fn one_op
|
|
(doc "One bench operation: build+sum a fresh CHUNK_LEN-cell list, pin the tree's root, return their sum so the value chain stays observable.")
|
|
(type
|
|
(fn-type
|
|
(params (con Int) (con Tree))
|
|
(ret (con Int))))
|
|
(params chunk_len t)
|
|
(body
|
|
(app + (app sum_list (app cons_n chunk_len)) (app pin_root t))))
|
|
|
|
; ---------- Bench loop ----------
|
|
|
|
; Loop runs `remaining` ops. Every PRINT_K ops, prints the
|
|
; rolling sum from the most-recent op (always equal to
|
|
; CHUNK_LEN*(CHUNK_LEN-1)/2 + 1 = 124750 + 1 = 124751 for
|
|
; CHUNK_LEN=500). The print is the timing event. The
|
|
; print_every counter's role is to keep stdout lines per second
|
|
; tractable for the harness (1000 timings instead of 20000).
|
|
;
|
|
; The tree `t` is passed through every recursive call so it
|
|
; stays a live root; Boehm has to trace through it on every
|
|
; collection.
|
|
|
|
(fn loop
|
|
(doc "Tail-recursive bench loop. Ops countdown in `remaining`; print marker every time `print_countdown` hits 0.")
|
|
(type
|
|
(fn-type
|
|
(params (con Int) (con Int) (con Int) (con Int) (con Tree))
|
|
(ret (con Unit))
|
|
(effects IO)))
|
|
(params remaining print_countdown chunk_len print_k t)
|
|
(body
|
|
(if (app == remaining 0)
|
|
(do io/print_int 9999)
|
|
(if (app == print_countdown 0)
|
|
(seq
|
|
(do io/print_int (app one_op chunk_len t))
|
|
(tail-app loop
|
|
(app - remaining 1)
|
|
(app - print_k 1)
|
|
chunk_len
|
|
print_k
|
|
t))
|
|
(let _v (app one_op chunk_len t)
|
|
(tail-app loop
|
|
(app - remaining 1)
|
|
(app - print_countdown 1)
|
|
chunk_len
|
|
print_k
|
|
t))))))
|
|
|
|
(fn main
|
|
(doc "Top-level: build tree, signal READY (8888), run loop, signal DONE (9999 emitted by loop).")
|
|
(type (fn-type (params) (ret (con Unit)) (effects IO)))
|
|
(params)
|
|
(body
|
|
(let t (app build_tree 19)
|
|
(let _root (app pin_root t)
|
|
(seq
|
|
(do io/print_int 8888)
|
|
(app loop 20000 0 500 20 t)))))))
|