メニュー
Use this skill to generate benchmark datasets (TPC-H, TPC-DS, etc.). Trigger when the user needs test data at a specific scale factor for benchmarking or testing. Supports parquet and duckdb output formats.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Run TPC-H benchmarks on Super Sirius or DuckDB CPU baseline — generate data, execute queries, validate results, and compare timings. Trigger when the user mentions benchmarking, TPC-H, performance testing, query runtimes, or wants to compare Sirius vs DuckDB speed.
Use this skill to analyze Sirius query execution logs — to find why a query is slow, diagnose validation errors (wrong results), spot excessive downgrading, identify pipeline gaps, or compare two runs of the same query. Trigger whenever the user mentions a Sirius log file, says "look at this log", asks about per-operator time, pipeline gaps, memory pressure, downgrade behavior, validation/correctness mismatches, or wants to compare two query runs. Also use as a hand-off target from race-check (for two-run comparisons), runtime-errors (for "where did the query stall before hang"), and validate (for diffing the bad run against a good run). This is the log-only counterpart to profile-analyzer (nsys) and optimization-advisor (source-level guidance).
Use this skill to find exactly which source code to optimize for better GPU performance. Maps nsys profile hotspots to specific Sirius source files and functions, classifies bottlenecks as GPU-bound, CPU-bound, or sync-bound, and recommends actionable code changes. Trigger when the user wants to know what to optimize, where to focus coding effort, or wants source-level optimization guidance. This skill focuses on actionable source code targets — for generating performance reports and measurements, use profile-analyzer instead.
Use this skill to understand why a Sirius query is slow, identify GPU bottlenecks, or detect performance regressions. Generates reports with kernel occupancy, memory bandwidth, operator attribution, and cross-run comparisons. Trigger when the user mentions profiling, nsys, GPU utilization, kernel analysis, performance reports, or wants to compare query timings across runs. This skill focuses on measurement and reporting — for mapping hotspots to source code fixes, use optimization-advisor instead.
Use this skill when the build fails, compilation errors occur, or you see undefined references, linker errors, CUDA compilation issues, missing headers, or template instantiation failures. Analyzes errors, suggests fixes, and iteratively rebuilds until success.
Discover and document a dependency library or submodule — analyzes all uses within the codebase, divides the library into logical modules, identifies which modules are used, and generates LLM-consumable API documentation for each module. Use when the user wants to understand a library dependency, map its modules, or generate API reference docs for a submodule.
| name | dataset-manager |
| description | Use this skill to generate benchmark datasets (TPC-H, TPC-DS, etc.). Trigger when the user needs test data at a specific scale factor for benchmarking or testing. Supports parquet and duckdb output formats. |
| argument-hint | [benchmark] [scale_factor] [--format duckdb|parquet] [--cluster] [--output <path>] |
Generate benchmark datasets by running the corresponding shell script under test/<benchmark>_performance/.
Parse $ARGUMENTS for:
--format duckdb|parquet (optional — each benchmark has a default)--cluster / --cluster-keys <spec> (optional — physically sorts tables so row-group pruning works; errors with --format parquet)--output <path> (optional — scripts have sensible defaults)If any required parameter (benchmark, scale factor) is missing, ask the user.
Each entry follows the same structure: script location, command template, supported formats, defaults, and prerequisites.
| Field | Value |
|---|---|
| Script | test/tpch_performance/generate_tpch_data.sh |
| Default format | parquet |
| Formats | parquet (tpchgen-rs), duckdb (DuckDB dbgen()) |
| Default output (parquet) | test_datasets/tpch_parquet_sf<SF> |
| Default output (duckdb) | test_datasets/tpch_sf<SF>.duckdb |
Default output (duckdb + --cluster) | test_datasets/tpch_sf<SF>_sorted.duckdb |
| Prerequisites | Parquet: pixi env (rust, python, pyarrow). DuckDB: build/release/duckdb |
cd test/tpch_performance && pixi run bash generate_tpch_data.sh <SF> --format <FORMAT> [--cluster] [--cluster-keys <spec>] [--output <path>]
Notes:
--cluster (duckdb only) physically sorts tables at load so each row group's per-column min/max becomes selective — this is what makes Sirius's native-scan row-group pruning skip work on date-filtered queries. It writes a distinct tpch_sf<SF>_sorted.duckdb so it can coexist with the unsorted dataset.lineitem:l_shipdate,orders:o_orderdate. Override with --cluster-keys "table:col,table:col,..." (implies --cluster).--cluster errors if combined with --format parquet (tpchgen-rs writes decimals as FIXED_LEN_BYTE_ARRAY, which disables Sirius row-group pruning for the whole file).| Field | Value |
|---|---|
| Script | test/tpcds_performance/generate_tpcds_data.sh |
| Default format | duckdb |
| Formats | duckdb, parquet |
| Default output (duckdb) | test_datasets/tpcds_sf<SF>.duckdb |
| Default output (parquet) | test_datasets/tpcds_parquet_sf<SF> |
| Prerequisites | build/release/duckdb |
cd test/tpcds_performance && bash generate_tpcds_data.sh <SF> --format <FORMAT> [--output <path>]
Notes:
test/tpcds_performance/queries/q{1..99}.sqlFor any benchmark that requires the DuckDB binary, check before running:
test -x build/release/duckdb
If missing, tell the user to build first: CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) make