Do not zero-fill the kvikio_ofstream staging buffer

[bdice]

kvikio_ofstream allocated its staging buffer (32 MiB by default) as a std::vector<char>, which value-initializes it, so every open memset the whole buffer. With glibc a 32 MiB request is always a fresh mmap, so the memset faults in all 8,192 pages: about 10 ms per index save, even when the file only receives a few KiB of header and the bulk data goes straight to kvikio (write_device, or host blocks at least as large as the buffer).

This PR allocates the buffer with std::make_unique_for_overwrite<char[]> and keeps its size in a member. Only cpp/src/util/file_io.cpp changes; the public API is unchanged.

• Same output. Only [pbase(), pptr()) is ever handed to kvikio, and xsputn / overflow write those bytes before advancing pptr(), so the bytes written are identical. A differential test of the old and new sbuf produced byte-identical files, including with the new buffer pre-filled with garbage under ASan/UBSan.
• Same errors. Open and allocation failures still raise Cannot open file ... for writing.
• Never slower. When the whole buffer is used, both versions fault the same pages.

A CPU microbenchmark of a 32 MiB allocation with a 4 KiB payload goes from ~10–12 ms to ~0.005 ms per open (8,193 → 2 minor page faults).

Testing

UTIL_TEST (including the FileIO.KvikioOfstream* tests) and the full ctest suite pass.

Measurements

Single-process wall time of each test executable on an RTX 6000 Ada (48 GB), otherwise idle (no ctest parallelism). The old and new libcuvs.so were run alternately, 2 or more repetitions each, with the order reversed between repetitions; the table shows the mean. The two builds differ only by this change and the test binaries are identical. All tests passed in every run.

executable	before	after	change
NEIGHBORS_ANN_VAMANA_TEST	316.3 s	283.8 s	-10.3%
NEIGHBORS_ANN_CAGRA_FLOAT_UINT32_TEST	88.5 s	84.2 s	-4.9%
NEIGHBORS_ANN_CAGRA_HALF_UINT32_TEST	51.5 s	47.9 s	-7.0%
NEIGHBORS_ANN_CAGRA_INT8_UINT32_TEST	74.5 s	71.6 s	-3.8%
NEIGHBORS_ANN_CAGRA_UINT8_UINT32_TEST	79.3 s	74.3 s	-6.2%
NEIGHBORS_ANN_HNSW_ACE_FLOAT_UINT32_TEST	7.8 s	7.6 s	-3.3%
NEIGHBORS_ANN_HNSW_ACE_HALF_UINT32_TEST	10.5 s	10.2 s	-2.8%
NEIGHBORS_ANN_HNSW_ACE_INT8_UINT32_TEST	8.6 s	7.4 s	-13.4%
NEIGHBORS_ANN_HNSW_ACE_UINT8_UINT32_TEST	7.6 s	7.6 s	-0.7%
NEIGHBORS_ANN_CAGRA_BBQ_UINT32_TEST (not affected)	3.8 s	3.9 s	+2.4%
NEIGHBORS_ANN_BRUTE_FORCE_TEST	5.1 s	4.0 s	-20.2%
NEIGHBORS_ANN_IVF_SQ_TEST	23.4 s	22.3 s	-5.0%
NEIGHBORS_ANN_IVF_RABITQ_TEST	11.8 s	10.4 s	-12.1%
UTIL_TEST	1.4 s	1.4 s	-2.5%
total	690.2 s	636.5 s	-7.8%

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[julianmi]

Good catch, @bdice. LGTM.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/cuvs/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 889b16c3-a0e0-4292-84dc-bb77e907a077

📥 Commits

Reviewing files that changed from the base of the PR and between 7f21ade and a7557dc.

📒 Files selected for processing (1)

• cpp/src/util/file_io.cpp

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 8 remain after this review.

---

📝 Summary

Summary by CodeRabbit

• Performance
  • File output now manages temporary write buffers more efficiently, including for larger writes. Existing file-writing behavior remains unchanged.

Walkthrough

The stream buffer now allocates staging storage with std::make_unique_for_overwrite and tracks its capacity separately. Initialization, direct-write checks, and flush resets use the tracked capacity and allocated pointer.

Changes

Stream Buffer Allocation

Layer / File(s)	Summary
Allocate and use the staging buffer cpp/src/util/file_io.cpp	The staging storage changes from std::vector<char> to an uninitialized std::unique_ptr<char[]> buffer with a separate capacity. Setup, direct-write checks, and flush resets use that capacity and the allocated pointer.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Refactor

Suggested reviewers: julianmi

Merge Risk: ⚪ Minimal · up to a7557

The change avoids zero-initializing the staging buffer while preserving the described write behavior. No merge-blocking issue was established; the PR is mergeable after normal checks.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 3 functions across 1 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: removing unnecessary zero-filling of the kvikio_ofstream staging buffer.
Description check	✅ Passed	The description directly explains the staging-buffer change, its performance impact, compatibility, and test results.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}