← Back

Performance Engineering

High-performance computing applied to cryptography: zkVM prover throughput, SIMD kernels, and sumcheck verifier optimization.

Valida zkVM — prover performance

Cryptography Engineer — Lita

January 2025 – 2025 · Fully remote

The reference workload was valida run --fast on reva-client-eth against Ethereum mainnet block 17034871. The flamegraph showed most time in the multi-segment proving path spent on avoidable clones — ValidaSegmentBootData::clone, <BTreeMap as Drop>::drop, and the inner BasicMachine::run / suspend loop.

Memory-aware segment splitting. Replaced the original "any-trace-near-max-height" splitting rule with a memory-operation-count rule (memory ops dominate trace height, and the count is a static function of the opcode). This eliminated the full first-pass logging that was contributing 30+ minutes on the reference block, enabling a two-pass flow: fast pass to discover segment boundaries, then per-segment slow-mode re-execution for trace generation.
Boot-data plumbing. Teased apart which fields of ValidaSegmentBootData were required for correctness (memory backend, persistent-bus state) versus only for trace generation, killing avoidable clones on segment boundaries while preserving the persistent-bus consistency check.
Parallel proving — coordinator/worker design. Built out the parallel prover (coordinator.rs, worker.rs, parallel.rs) with four sync barriers (trace, permutation, quotient, opening). The coordinator owns the Fiat-Shamir transcript (transcript-soundness requires sequential challenge sampling); workers handle trace generation, polynomial ops, quotient and permutation construction. prove_multi_segment_keccak_parallel ran at ~38s wall / 167s CPU (4.3× speedup) with ~78 MB peak RSS.
STARK circuit performance. R&D of high-performance STARK circuits for the Valida chip set, including the degree-3-safe XOR3 expansion in the Keccak-f chip.
Profiling infrastructure. Set up cargo flamegraph with kernel.perf_event_paranoid and release-build debug symbols; iterated on flamegraphs of the multi-segment proving path.

WHIR / Plonky3 — SIMD and sumcheck

Merged SIMD compress_hi_dot PR in Plonky3. Rewrote the WHIR compress_hi_dot kernel to route the Packed arm through Plonky3's F::Packing::dot_product::<4> delayed-reduction SIMD dot-product. Two wins stack: ~4× fewer Montgomery reductions in the inner loop (one per 4-wide group instead of one per multiply), and instruction-level parallelism from four independent per-coefficient accumulators replacing the serial mul-add chain. Includes a narrow unsafe block using MaybeUninit::uninit() to skip eager zero-fill of a [F::Packing; 128] scratch buffer.
FRI vs WHIR Pareto frontier — benchmark PR. Extended a single-point FRI/WHIR comparison into a full parameter sweep tracing each scheme's prover-cost-vs-argument-size frontier on an identical claim — a univariate↔multilinear bridge opening 256 polynomials at one common point at 100-bit soundness. The WHIR side randomized-grid-searches per-round folding, starting inverse-rate, and per-round rate schedules; the FRI side sweeps log blowup factor and max arity, deriving query counts to hit the soundness target. Each configuration reports a theoretical proxy (total committed oracle length in base-field elements, extension-field oracles counted ×4), measured wall-clock prover time, and postcard argument size. The frontier shows WHIR sitting below FRI in theory, but FRI winning the cheap-prover regime on parallelism while WHIR reaches the smallest proofs at higher sequential sumcheck cost — with high folding inflating argument size, so moderate folding is the sweet spot.
"One sumcheck optimization a day" — Twitter thread series. Daily public-facing unpacking of Speeding Up Sum-Check Proving (Dao, DeStefano, Bagad, Domb, Thaler — ePrint 2026/587). Each post under 280 characters, tuned for the BabyBear/KoalaBear setting where whir-p3 lives.
Long-running Claude Code optimization agent. Built a long-running Claude Code agent setup (OPTIMIZATION_PLAN.md + context/ + findings/) that scans the Plonky3/whir/ codebase against the paper's full optimization taxonomy, scores each technique for applicability/composability/effort, and writes per-optimization findings — with hard architectural constraints (KoalaBear 31-bit base field, two sumcheck cost centers: OOD samples and STIR queries) baked into the prompt.

Toolkit

Rust — primary implementation language.
SIMD: Plonky3 packed-field intrinsics; delayed-reduction kernel design.
Multilinear sumcheck protocol; verifier-side optimization.
Profiling: perf, cargo flamegraph, Criterion benches.
Data science (Kaggle bronze medal — RSNA-MICCAI brain tumor classification, 2021).