Skip to content

bench

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

bench

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
loop_policy_aggregate.py
loop_policy_aggregate.py
 
 
loop_policy_sweep.sh
loop_policy_sweep.sh
 
 
sweep.sh
sweep.sh
 
 

README.md

ITB microBatch Policy Sweep

Single-file sweep runner for the CGO pixel encoder's adaptive microBatch switch and its two sync.Pool tiers. The shipped binary reads the policy from two env vars at package init(), so no source edits are needed between runs.

Env vars

Var Format Default (unset)
ITB_MICROBATCH_TIERS "upper:batch,…,-1:batch" 16384:512,8388608:262144,-1:512
ITB_HASHPOOL_STARTERS "s1,s2,…" (ints, uniq+sorted) 512,262144

ITB_MICROBATCH_TIERS

A comma-separated list of upper:batch pairs read left to right. The first tier whose upper matches (or the -1 fallback) picks the batch stride. upper is compared against the per-region payload length — under Triple Ouroboros the low-level entry receives one third of the user-visible plaintext.

Examples:

  • Shipped 3-tier: 16384:512,8388608:262144,-1:512
  • 4-tier with mid stride: 16384:512,131072:65536,8388608:262144,-1:512
  • 2-tier always adaptive: 16384:512,-1:262144
  • Baseline everywhere: -1:512

ITB_HASHPOOL_STARTERS

A comma-separated list of sync.Pool starter capacities (in uint64 elements). Each entry becomes an independent pool tier; every fresh pool.New call allocates two arrays of that size.

getHashArraysFor(microBatch, n) picks the smallest tier whose starter >= microBatch; falls back to the largest tier when microBatch exceeds every configured starter.

Examples:

  • Shipped 2-pool: 512,262144
  • 3-pool with mid tier: 512,65536,262144
  • Single pool baseline: 512
  • Single pool always wide: 262144

Compact output format

CPU: <model> <cores>c/<threads>t
Policy: <name>
  tiers=<active tiers or "default">
  pools=<active pools or "default">
E   4KB   <allocs/op>   <ns/op>   <MB/s>
E  64KB   <allocs/op>   <ns/op>   <MB/s>
…
D   4KB   <allocs/op>   <ns/op>   <MB/s>
…

E = Encrypt, D = Decrypt. Rows in ladder order across 4KB / 64KB / 512KB / 1MB / 4MB / 8MB / 16MB / 32MB / 48MB / 64MB.

Usage

The runner prints a header (CPU model, policy name, active tier / pool overrides) followed by the compact E / D rows across the size ladder. Set POLICY_NAME for the header label and any of ITB_MICROBATCH_TIERS / ITB_HASHPOOL_STARTERS for the tuner axes; unset knobs fall back to the shipped defaults.

A one-time shell prep line covers the toolchain env every example below assumes:

export PATH=/usr/local/go/bin:$PATH

Policy grid

Baseline — shipped defaults:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="baseline" \
    bash scripts/bench/sweep.sh

3-tier mid stride at 65 536 with a matching 3-pool:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="3tier-mid65536" \
    ITB_MICROBATCH_TIERS="16384:512,131072:65536,8388608:262144,-1:512" \
    ITB_HASHPOOL_STARTERS="512,65536,262144" \
    bash scripts/bench/sweep.sh

3-tier mid stride at 131 072 with a matching 3-pool:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="3tier-mid131072" \
    ITB_MICROBATCH_TIERS="16384:512,1048576:131072,8388608:262144,-1:512" \
    ITB_HASHPOOL_STARTERS="512,131072,262144" \
    bash scripts/bench/sweep.sh

Wider upper tier — raise the adaptive boundary to 16 MiB:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="wider-upper" \
    ITB_MICROBATCH_TIERS="16384:512,16777216:262144,-1:512" \
    bash scripts/bench/sweep.sh

Always adaptive above 16 KiB — one wide stride for everything larger:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="always-wide" \
    ITB_MICROBATCH_TIERS="16384:512,-1:262144" \
    bash scripts/bench/sweep.sh

Single-tier and single-pool baseline:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="single-tier" \
    ITB_MICROBATCH_TIERS="-1:512" \
    ITB_HASHPOOL_STARTERS="512" \
    bash scripts/bench/sweep.sh

Single large pool — isolate the effect of pool sizing:

CGO_ENABLED=1 ITB_INNER_HASH=areion512 POLICY_NAME="single-large-pool" \
    ITB_HASHPOOL_STARTERS="262144" \
    bash scripts/bench/sweep.sh

Additional isolation policies compose the same env vars in obvious ways — e.g. default+pool65536 keeps the shipped tier switch but adds the mid pool tier alone (ITB_HASHPOOL_STARTERS="512,65536,262144") to measure whether the pool starter moves anything on its own.

Env overrides

Var Default Notes
POLICY_NAME default Human label printed in the header
BENCH_TIME 1s go test -benchtime value
BENCH_COUNT 1 go test -count value (median of N)
ITB_REPO $PWD/../.. Repo root path
ITB_INNER_HASH areion512 Canonical bench profile knobs
ITB_KEY_BITS 512
ITB_NONCE_BITS 512
ITB_INTERLOCK "" (48) Interlocked Barrier chunk width (48 or 168; 168 requires a 512-bit ITB_INNER_HASH)
ITB_WITH_MAC false
ITB_WITH_PARALLAX false
ITB_WITH_WRAPPER false
ITB_GOMEMLIMIT 1GiB
ITB_GOGC 20

Wall-time budget

1s × count=1 × 20 sub-benches ≈ 20 s per policy on a warm cgo cache plus ~5 s of pre-run cgo build. Cold cache adds ~30–60 s to the first policy of a session.

If a policy shows visibly noisy numbers, bump BENCH_TIME=2s; if still noisy on the very small end of the ladder, bump BENCH_COUNT=3 and read the median across runs.

Steady-state policy sweep via tools/loop

loop_policy_sweep.sh drives the same two policy env vars through tools/loop instead of go test -bench: one long-lived Streaming AEAD pipeline per cell, three worker goroutines, a fixed run duration, and the loop's --json-output summary captured per cell. The loop echoes the active policy on a policy: log line and in the microbatch_tiers / hashpool_starters JSON fields, so every cell is self-describing.

Cells are listed one per line in a pipe-delimited spec file (label|tiers|pools|gogc|hash|payload|extra loop flags; - selects the default for that column, # starts a comment):

bash scripts/bench/loop_policy_sweep.sh cells.txt /tmp/policy-out
python3 scripts/bench/loop_policy_aggregate.py /tmp/policy-out --sort payload

LOOP_DURATION (default 90s), LOOP_MEMLIMIT (default 2GiB, applied both as GOMEMLIMIT and --memlimit), LOOP_GOGC (default 85), LOOP_HASH (default aesitb128) and LOOP_COMMON (the fixed loop flags — Streaming AEAD with hmac-blake3, parallax and wrapper on, 1024-bit key, three goroutines) override the fixed knobs; LOOP_BIN points at a prebuilt loop binary instead of building one into the output directory.

The aggregator prints a Markdown table with the policy knobs, encrypt / decrypt throughput, allocation per iteration, GC rate and cost, and the per-pool miss rates. The figures are steady-state Streaming AEAD numbers and are not comparable with the cold-start Single Message rows sweep.sh produces; the tier thresholds are per-region byte counts, so under the default 16 MiB chunk budget a stream never reaches the shipped ladder's third tier — probing it takes an explicit --chunk-size above three times the second tier's upper bound.