feat(schema): optional GPU device-memory fields (gpu_allocations_bytes, gpu_live_bytes) — Piccolissimo #252 - #17
Open
jeonghun-jj-lee wants to merge 1 commit into
Open
Conversation
…iccolissimo #252)
Add two optional, defaulted fields to the BenchmarkResult schema so the GPU
arms of the dimension-ladder scaling benchmark (Piccolissimo #249, ADR 0007)
can record device memory — the axis the benchmark exists to measure:
- gpu_allocations_bytes: cumulative device allocation over the timed solve
- gpu_live_bytes: post-solve live device bytes
Both are Union{Nothing,Int} defaulting to `nothing`. Absent-value semantics:
`nothing` = "device axis not measured" (CPU-arm solves, and any result
serialized before these fields existed); a concrete `0` is distinct and means
the device WAS measured with a zero delta. Field names mirror the host
`total_allocations_bytes` convention. Capturing the values during solves is
out of scope here (consumed by the GPU-arms slice, Piccolissimo #254).
Backward-compat (the hard invariant): a BenchmarkResult JLD2-serialized under
the old schema deserializes as a JLD2.ReconstructedMutable once the struct
gains fields, which load_results could not convert. Added a
`convert(::Type{BenchmarkResult}, ::JLD2.ReconstructedMutable{:BenchmarkResult})`
upgrade shim in storage.jl that reads the fields the old blob carries and
defaults any it lacks (written generically so it also tolerates future
additive changes). A committed pre-change fixture
(test/fixtures/pre_gpu_fields_v1_pregpu0.jld2) plus a regression test prove
old baselines still load with the GPU fields defaulting to `nothing`.
Comparison path is untouched: no comparison keys on the new fields (AC3).
Tests: 333 pass, including new gpu-fields default/round-trip testsets and the
backward-compat fixture load.
Refs: Piccolissimo #252, #249; ADR 0007 (PR #250).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Adds two optional, defaulted device-memory fields to the
BenchmarkResultschema so the GPU arms of the dimension-ladder scaling benchmark (Piccolissimo #249, ADR 0007 / PR #250) can record the axis the benchmark exists to measure:gpu_allocations_bytes— cumulative device allocation over the timed solve (deterministic GPU analogue of hosttotal_allocations_bytes).gpu_live_bytes— post-solve live device bytes.Both are
Union{Nothing,Int}defaulting tonothing.Absent-value semantics (documented on the schema):
nothing= "device axis not measured" — CPU-arm solves, and any result serialized before these fields existed, deserialize asnothing. A concrete0is distinct: it means the device WAS measured and the delta was zero.Capturing the values during solves is out of scope here — that is consumed by the GPU-arms slice (Piccolissimo #254). This PR is the schema + serialization contract only.
Backward-compatibility guarantee (the hard invariant)
A
BenchmarkResultJLD2-serialized under the old schema deserializes as aJLD2.ReconstructedMutable{:BenchmarkResult}once the struct gains fields — whichload_resultscould notconvertintoVector{BenchmarkResult}(it threw aMethodError). This PR adds aconvert(::Type{BenchmarkResult}, ::JLD2.ReconstructedMutable{:BenchmarkResult})upgrade shim instorage.jlthat reads the fields the old blob carries and defaults any it lacks. It is written generically over the reconstructed field set, so it also tolerates future additive schema changes.Proof: a committed pre-change fixture
test/fixtures/pre_gpu_fields_v1_pregpu0.jld2(serialized with the schema before these fields existed) plus a regression test that loads it and asserts it comes back as a realVector{BenchmarkResult}with the GPU fields defaulting tonothingand every pre-existing field intact.Comparison path
Untouched — no comparison keys on the new fields (AC3). CPU-arm results are unaffected (fields simply default).
Tests
Pkg.test()→ 333 pass, including:GPU device-memory fields (schema + defaults)— omitted ⇒nothing; explicit0distinct fromnothing; real GPU values carried.GPU device-memory fields: JLD2 round-trip— new results with GPU values and with defaulted (nothing) fields both round-trip.Backward-compat: pre-GPU-fields fixture still loads— the committed pre-change fixture loads with fields defaulted.Refs
Piccolissimo #252 (parent #249); ADR 0007 (PR #250). Blocks Piccolissimo #254. After merge, the Piccolissimo scaling env gets re-pinned to this revision (handled separately, per SHA-pinned-env convention).
🤖 Generated with Claude Code