Thaw

perf: parallel CRC32C — restore 2.89× faster

This release lands a two-edit perf pass on crates/thaw-runtime/src/pipeline.rs that makes the pipelined restore path CRC-parallel and decouples the CRC fold from the pread path. CRC content integrity is still fully verified; THAW_VERIFY=0 still bypasses.

Throughput (16 GiB synthetic, 1× H100 SXM, warm page cache)

| Operation | Before | After | Speedup |
|---|---|---|---|
| restore_from_file_pipelined (chunk=64) | 4.80 GB/s | 13.88 GB/s | 2.89× |
| restore_from_file_pipelined (chunk=256) | 4.51 GB/s | 14.56 GB/s | 3.23× |
| restore_from_file_pipelined (chunk=1024) | — | 15.47 GB/s | — |
| freeze_to_file_pipelined (chunk=64) | 3.32 GB/s | 3.88 GB/s | 1.17× (NVMe-bound) |
| restore_from_bytes_pipelined (mmap) | 1.43 GB/s | 1.64 GB/s | 1.15× (PTE-walk-bound) |

What changed

• New helper crc32c_append_parallel — splits slices ≥4 MiB into 8 shards, computes crc32c on each via std::thread::scope, then stitches with crc32c_combine. Serial fallback for smaller slices avoids spawn overhead.
• Steady-state loop in restore_pipelined reordered — launch_uploads(prev) fires first; the CRC fold for the previous chunk runs on a worker thread in parallel with read_chunk(curr) on the main thread. Critical path becomes max(CRC, pread) instead of CRC + pread.
• Call-site swap — accumulate_freeze_crcs, fold_and_verify_chunk_crc, and verify_plan_crcs_from_bytes all use the parallel CRC, so freeze, pipelined restore, and in-memory restore all benefit.

No public API change.

Correctness

• All 76 thaw-runtime unit tests pass, including:
  • restore_detects_single_byte_payload_corruption
  • pipelined_from_bytes_detects_payload_corruption
  • restore_round_trips_through_freeze
• Manual bit-identity round-trip on 2 GiB of f32 random data: pass.

Caveats

• Warm-path measurement. This pod has 2 TB RAM; the 16 GiB file lives in page cache after freeze. Cold O_DIRECT throughput on this pod caps at ~2.5 GB/s single-stream / ~8.4 GB/s parallel 4× — pod-specific, not code-bound.
• Freeze bump is small because NVMe write is the ceiling on this pod (dd caps at 3.1 GB/s).

Install

pip install --upgrade thaw-vllm thaw-native

• thaw-native 0.3.2 (Rust runtime — the perf change lives here)
• thaw-vllm 0.5.1 (Python integration — pin bump + version sync)

Reproduce

python3 benchmarks/micro_pipeline.py --size-gb 16 --chunk-mb 256 --repeats 3 --path /tmp/micro.thaw

Still on the table

• Persistent CRC worker (one spawn per restore instead of one per chunk).
• Multi-queue pread ring (unlocks parallel NVMe reads on cold path).
• AVX-512/PCLMUL CRC32C (crates that double single-core CRC).