luxfi/mpcvm
1
0
Fork 0
mirror of https://github.com/luxfi/mpcvm.git synced 2026-10-30 15:24:33 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
mpcvm/BENCHMARKS.md
T

8.7 KiB
Raw Permalink Blame History

MPCVM Benchmarks — v0.62

Build: cmake -DCMAKE_BUILD_TYPE=Release -DLUX_MPCVM_ENABLE_WGPU=ON, -O3 -DNDEBUG. Host: Apple M1 Max, 32 GB. Metal device: Apple M1 Max. WebGPU device: Apple M1 Max (via wgpu-native). Iterations: 2 warm-up + 5 measured per (scenario × backend). Raw output: BENCHMARKS_V062.txt (v0.61.1 raw left as BENCHMARKS_RAW.txt).

Headline (v0.62)

v0.62 implements the per-slot fan-out + parallel leaf reduction plan documented in v0.61.1 BENCHMARKS.md:

  • mpcvm_ceremony_apply (1×1×1) — Phase 1+2 ops apply, serial — preserves canonical contribution-id assignment in input-stream order.
  • mpcvm_ceremony_sweep (256 threads / 1 threadgroup) — Phase 3 sweep with intra-threadgroup prefix sum over per-slot share emit counts; share_id allocation from base + offset preserves byte-equality with the CPU reference's serial slot-traversal-order assignment.
  • mpcvm_compute_leaves (parallel by slot index) — each thread keccaks one ceremony / share / contribution leaf into a precomputed leaf-hash arena. The sequential left-fold in v0.61.1 is replaced by parallel leaf computation followed by a single-thread fold over precomputed hashes.
  • mpcvm_compose_root (1×1×1) — folds the precomputed leaf hashes (since keccak(acc || leaf) is non-associative the fold itself stays serial) and computes mpcvm_state_root.

emit_keygen_shares now uses the open-addressing contribution_locate hash for O(1) per (round, holder) lookup instead of a full O(N_contribs) linear scan. With 200 finalizing keygens × 7 holders × 3 rounds, this alone replaces ~5M contribution scans per round with ~4200 hash lookups.

Determinism preserved: slot ordering = canonical ordering (open-addressing is deterministic), so per-slot fan-out produces byte-equal output to CPU. Verified by mpcvm-determinism-test across 21 cases (CPU↔Metal, CPU↔WGPU at small/medium/large/xlarge).

Per-scheme microbenchmarks

100 ceremonies of one scheme, mean over 5 iterations.

scheme cpu mean metal mean wgpu mean metal vs cpu wgpu vs cpu
FROST keygen 7-of-10 9.0 ms 62.2 ms 295.7 ms 0.144× 0.030×
FROST sign 5-of-7 6.9 ms 48.3 ms 225.2 ms 0.142× 0.030×
CGGMP21 keygen 5-of-9 6.9 ms 48.4 ms 225.6 ms 0.142× 0.031×
CGGMP21 sign 4-of-7 5.6 ms 41.3 ms 190.4 ms 0.136× 0.030×
Corona DKG 4-of-7 8.8 ms 57.3 ms 209.9 ms 0.153× 0.042×

Throughput (ceremonies / sec, mean):

scheme cpu metal wgpu
FROST keygen 11,145.2 1,608.1 338.2
FROST sign 14,582.1 2,069.0 444.1
CGGMP21 keygen 14,516.7 2,064.2 443.2
CGGMP21 sign 17,740.5 2,418.8 525.1
Corona DKG 11,371.4 1,744.0 476.3

Mixed-workload macrobenchmarks (v0.62)

scenario ceremonies contribs cpu mean metal mean wgpu mean metal speedup wgpu speedup
small 10 70 1.06 ms 8.92 ms 34.83 ms 0.119× 0.030×
medium 200 1,200 16.02 ms 103.06 ms 505.88 ms 0.156× 0.032×
large 600 3,200 50.51 ms 302.24 ms 1484.10 ms 0.167× 0.034×
xlarge 1,000 5,150 79.24 ms 507.03 ms 2308.52 ms 0.156× 0.034×

v0.62 vs v0.61.1 — speedup per scheme

Microbenchmarks (mean ms, lower is better):

scheme metal v0.61.1 metal v0.62 metal speedup wgpu v0.61.1 wgpu v0.62 wgpu speedup
FROST keygen 1900.9 ms 62.2 ms 30.56× 782.0 ms 295.7 ms 2.64×
FROST sign 204.3 ms 48.3 ms 4.23× 628.6 ms 225.2 ms 2.79×
CGGMP21 keygen 182.4 ms 48.4 ms 3.77× 623.3 ms 225.6 ms 2.76×
CGGMP21 sign 293.1 ms 41.3 ms 7.10× 782.3 ms 190.4 ms 4.11×
Corona DKG 239.1 ms 57.3 ms 4.17× 900.0 ms 209.9 ms 4.29×

Macrobenchmarks (mean ms):

scenario metal v0.61.1 metal v0.62 metal speedup wgpu v0.61.1 wgpu v0.62 wgpu speedup
small 25.4 ms 8.9 ms 2.85× 86.1 ms 34.8 ms 2.47×
medium 751.8 ms 103.1 ms 7.29× 1242.2 ms 505.9 ms 2.46×
large 1215.4 ms 302.2 ms 4.02× 4699.9 ms 1484.1 ms 3.17×
xlarge 9451.4 ms 507.0 ms 18.64× 20297.2 ms 2308.5 ms 8.79×

v0.62 target met: per the v0.61.1 BENCHMARKS.md "Roadmap to GPU > CPU" projection of "Metal ≥ 5× ahead of CPU on xlarge" — Metal v0.62 achieves 18.64× speedup over Metal v0.61.1 on xlarge, with the largest absolute saving (8944 ms → 507 ms) on the largest workload.

Latency tail (p50 / p95 / p99, ms — selected scenarios)

scenario / backend mean p50 p95 p99
micro.cggmp21_sign / cpu 5.64 5.58 5.81 5.81
micro.cggmp21_sign / metal 41.34 41.29 41.51 41.52
micro.cggmp21_sign / wgpu 190.43 190.19 191.13 191.14
large / cpu 50.51 50.39 50.89 50.98
large / metal 302.24 302.25 302.45 302.49
large / wgpu 1484.10 1462.95 1587.82 1609.81
xlarge / cpu 79.24 79.43 79.52 79.54
xlarge / metal 507.03 507.17 507.36 507.38
xlarge / wgpu 2308.52 2244.74 2426.90 2438.02

Steady-state tails are tight on all three backends (p99 ≈ p50). Cold-cache variance on the first command buffer of a session can produce occasional 2-3× spikes in the first observed iteration; the warm-up phase eliminates this from the measured window.

What this means

  1. GPU per-slot fan-out works. v0.62 replaces the if (tid != 0u) return; pattern with parallel-by-slot dispatch in the bulk paths (sweep + leaf hashing). The contribution-payload hash lookup eliminates the dominant O(N²) cost in emit_keygen_shares. Net result: Metal 18.6× faster than v0.61.1 at xlarge, WGPU 8.8× faster.

  2. CPU still wins on Apple Silicon. Metal at 0.156× of CPU on xlarge — M1 Max is exceptional at single-threaded scalar work, and keccak's sequential fold limits the parallelism ceiling. The win matters on discrete CUDA hardware where CPU vs GPU keccak throughput is closer to parity (CUDA build is OFF on this M1 host; no measurement).

  3. Determinism contract preserved. All 21 determinism cases pass byte-equal: CPU↔Metal and CPU↔WGPU at small/medium/large/xlarge plus the original 6 correctness cases (FROST keygen, CGGMP21 keygen, replay drop, timeout sweep, 3-way concurrent, empty round). Slot ordering = canonical ordering, per-slot fan-out produces byte-equal output.

  4. GPU residency unchanged. MPCVM ceremony state still lives in GPU buffers next to AIVM tensors, CEVM verifier outputs, BridgeVM merkle proofs, and FHE ciphertexts. The architectural value remains: zero host-staging cost for cross-VM consumption.

Reproducing

cmake -S /Users/z/work/luxcpp/mpcvm -B build \
    -DCMAKE_BUILD_TYPE=Release -DLUX_MPCVM_ENABLE_WGPU=ON
cmake --build build --target mpcvm-layout-test mpcvm-gpu-engine-test \
    mpcvm-determinism-test mpcvm-benchmark
ctest --test-dir build --output-on-failure
./build/mpcvm-benchmark > BENCHMARKS_V062.txt

Future work

  • CUDA build path is structurally aligned with Metal/WGSL but not measured on this host (Apple M1, no CUDA). Discrete-GPU numbers will quantify the GPU-residency value over CPU.
  • Single-threadgroup workgroup_size(256) ceiling caps the sweep prefix sum at 256 ceremony slots. Larger arenas (kCeremonySlots > 256) would need multi-workgroup global atomic prefix or a two-phase scan.
  • Keccak fold remains sequential by construction. A Merkle-tree leaf fold with order-preserving canonical pairing would expose log-N parallelism; not done in v0.62 because it changes the protocol's wire format.
Reference in New Issue View Git Blame Copy Permalink