| name | velox-flamegraph-analyzer-skill |
| description | Analyze async-profiler flame graph HTML files to identify Java, Velox C++, and kernel bottlenecks, map hotspots to Velox modules, provide ARM/aarch64 optimization insights, and generate an interactive HTML report with links to original flame graphs.
|
Velox Flame Graph Analyzer Skill
Analyze async-profiler flame graph HTML files to identify Java/C++(Velox)/kernel bottlenecks, map hotspots to Velox business code modules, provide ARM/aarch64-specific optimization insights for the Velox engine, and generate an interactive HTML summary report with direct links to original flame graphs.
When to Use
- User provides a directory of async-profiler flame graph HTML files and asks for bottleneck analysis
- User asks about performance optimization on ARM/Kunpeng/aarch64 platforms
- User wants a visual summary of flame graph findings with clickable links
- User needs to cross-reference flame graph hotspots with source code
How to Invoke
python scripts/analyze.py \
--input <path_to_flame_graph_html_directory> \
--output <path_to_output_directory> \
--arch aarch64 \
--engine "Gluten 1.3.0 + Velox" \
--flame-rel-path <relative_path_from_report_to_flamegraphs>
Parameters
| Parameter | Required | Default | Description |
|---|
--input | Yes | - | Directory containing async-profiler flame graph HTML files |
--output | No | . | Output directory for the HTML report |
--arch | No | aarch64 | Target architecture: x86, arm, or aarch64 |
--engine | No | Gluten + Velox | Engine name shown in report header |
--sample-size | No | 60 | Number of files to sample for analysis (for large collections) |
--flame-rel-path | No | "" | Relative path prefix for flame graph links in the report |
--report-name | No | flamegraph_analysis_report.html | Output report filename |
--biz-keywords | No | Gluten+Velox set | Comma-separated keywords to identify business logic frames (vs runtime frames like libjvm.so) |
What the Tool Does
- Samples flame graph files from the input directory (stratified by time batch)
- Stream-parses each HTML file, extracting
f() call data from async-profiler JS
- Builds hotspots: computes self-time percentage for each stack frame relative to root
- Classifies frames into Java, C++/native, kernel, or other categories
- Identifies business code using configurable keyword matching (default: Velox/Gluten)
- Matches ARM rules: checks each business hotspot against known ARM performance patterns
- Builds conclusions grouped by functional area (Scan/IO, Join, Aggregate, Expression, Shuffle, Sort, Memory, JNI) with severity ratings
- Generates HTML report with:
- Summary dashboard (Java%, C++%, Kernel%, Business Code%)
- Java bottleneck analysis with per-area breakdown and optimization tips
- C++/native bottleneck analysis with ARM-specific notes and tags
- Business code top 30 hotspot table with module classification
- Kernel bottleneck table
- Representative flame graph file listings
- Overall ranked bottleneck summary table
- ARM/aarch64 platform-specific notes section (when arch=arm/aarch64)
- Every hotspot row includes a direct link to open the corresponding flame graph
ARM-Specific Analysis Rules
The tool recognizes these ARM/aarch64 performance patterns and provides targeted advice:
| Pattern | ARM Issue | Suggestion |
|---|
RleDecoderV2 | RLE/BIT_PACK lacks NEON vectorization | Add NEON decode paths; -march=armv8.2-a+dotprod |
HashStringAllocator | Frequent small malloc/free penalized by ARM cache | Switch to jemalloc+large pages; pre-allocate in bulk |
hashBytes/Murmur3Hash | Hash may not use ARM CRC32/crypto extensions | -march=armv8.2-a+crypto for CRC32 instructions |
HashTable/groupProbe | Hash probe suffers higher ARM memory latency | Use prefetch hints; larger initial allocation |
simdjson/GetJsonObject | get_json_object is scalar per-row; no vectorization | Cache JSON results; use batch Velox UDF |
CollectList/deserializeToValueList | UnsafeRow deserialization heavy on ARM string copy | Optimize with bulk memcpy; NEON string copy |
Custom Business Keywords
The default keyword set covers Gluten + Velox. For other engines, pass custom keywords:
python scripts/analyze.py --input ./flames --biz-keywords "ClickHouse,CH,clickhouse,NativeExecutor"
python scripts/analyze.py --input ./flames --biz-keywords "velox,flink,Arrow,substrait"
Output
A single HTML file (default: flamegraph_analysis_report.html) that can be opened in any browser. The report is self-contained (no external dependencies) and includes clickable links to open original flame graphs directly.
After Generating the Report
- Tell the user the report file path
- Summarize the top 3-5 key findings in plain text
- Highlight the most critical ARM-specific bottleneck if arch=aarch64
- Mention that flame graph links in the report are clickable and open the original async-profiler interactive visualization
