Skip to content

CLOUD-RUNBOOK.md

Latest commit

 

History

History
34 lines (28 loc) · 2.82 KB

CLOUD-RUNBOOK.md

File metadata and controls

34 lines (28 loc) · 2.82 KB
Raw

Runbook: native Linux on rented GPUs (planned)

Three rented machines answer the open questions in about an hour each. Anyone can run this; replication reports are welcome.

Machine Answers Rough cost
RTX 3090 (sm_86, 24 GB), native Linux Does the port build and prove on Ampere? Is patch 12 (WSL2 pinned memory) needed on native Linux? about $0.30/h
RTX 4090 (sm_89, 24 GB), native Linux The same for Ada, the consumer twin of the L4 about $0.50/h
H100 (sm_90a), native Linux, unmodified upstream Byte-for-byte check of the patches, including hiding (HIDING.md) about $2–3/h

Host requirements: a VM or bare-metal container with full syscalls. Avoid gVisor-style sandboxes, because cp_commit needs madvise(MADV_WIPEONFORK). You also need an x86-64 CPU with SHA extensions (grep -m1 -ow sha_ni /proc/cpuinfo), at least 10 GB of free RAM, Ubuntu 24.04, and g++ 13.

A. Consumer cards (3090 / 4090)

For sm_86 or sm_89, change patch 1's flag (-arch=sm_120) to the card's architecture; everything else is the same.

  1. Follow the README's "Reproduce it" steps 1–5, first with patches 1–11 only (without patch 12).
  2. If cp_commit fails at hf_open_side.cuh with invalid argument, apply patch 12 and record that it is needed on native Linux too.
  3. Run run_target_example.sh target_small_f2_roomy three times; check each with scripts/a5_verify_and_tamper.sh.

B. Byte identity on H100 (unmodified upstream vs our patches)

Goal: identical bytes from identical inputs under a fixed seed.

  1. Determinism here: done. Two test-build runs with the same seeds were byte-identical on the RTX 5070 Ti (scripts/a15_determinism.sh).
    • Seed files: 32 bytes each. cp_commit -test-seed: the ASCII text why-research A15 fixed test seed!. cp_prove -test-coins: why-research A15 fixed testcoins.
    • Configuration and request: target_small_f2_roomy with made-up-blocks, -M 1, and the request own_choices.req.
    • Reference: commitment ID 4942be953206913cf3c97c443df610f7ed9c03c0bc1b99437c17c2a11095cfb6
      • ivk/key.bin sha256 69f94db22375a92531466aa8160b5cf33fed390ac6178c5889c244603d770751
      • ivk/leaf/0.bin sha256 64a4fb799ccd54c5674c70688102e64cf50e8f730b21b8c8a630fc6b1a3bba14
      • out/example.proof sha256 152b7bf02689e62f9639fd667a8efa8babcab59a07efe9b9286cbf7e8eec3e77
      • out/example.y sha256 871cd8edd0dbd8de891216d6409fba14bb5f0f72b880c5fcc512218b282298fb (token 284)
  2. On the H100: build unmodified upstream 9e2bdd8f (./build.sh tests), run the same seed and inputs, and sha256sum key, leaf and proof.
  3. Compare with the hashes from this card. Equal hashes mean the 12 patches change no output byte, masks included.

Report: GPU, driver, OS, exit codes, last lines, IDs, and the sha256 of each output file.