Execution optimizations and improvements#179
Open
anubhavchaturvedi wants to merge 6 commits intometa-pytorch:mainfrom
Open
Execution optimizations and improvements#179anubhavchaturvedi wants to merge 6 commits intometa-pytorch:mainfrom
anubhavchaturvedi wants to merge 6 commits intometa-pytorch:mainfrom
Conversation
Adds a standalone benchmark (benches/parse_benchmark.rs) that measures: - Wall time statistics (mean/median/min/max) across configurable iterations - Peak RSS via getrusage (cold-run and post-warmup measurements) - Input file line count for context Usage: TLPARSE_BENCH_INPUT=/path/to/log cargo bench --bench parse_benchmark No production code changes. Dev-dependencies added: libc (RSS), tempfile (output dirs). - Removed hardcoded machine-specific path; requires explicit input - Added cold-run RSS measurement with documentation of ru_maxrss limitations - Streaming line count instead of loading entire file into memory - Write errors surfaced via expect() instead of silently swallowed
… year caching Four localized optimizations with zero API changes: 1. Pre-allocate HTML string in anchor_source (parsers.rs) - Remove intermediate Vec<&str> from lines().collect(), iterate directly - Pre-allocate output with String::with_capacity(text.len() * 2 + 500) 2. Pre-allocate shortraw_content buffer (lib.rs) - Use String::with_capacity(file_size / 8) (~12.5% of input size) - Avoids ~30 reallocations for large logs 3. Reuse payload String across parse loop iterations (lib.rs) - Hoist payload_buf before loop, clear() each iteration - Retains allocated capacity, avoiding millions of small allocations 4. Compute year once before parse loop (lib.rs) - Move chrono::Utc::now().year() before format_timestamp closure - Eliminates one clock_gettime syscall per log line Note: syntect lazy-init (SyntaxSet/ThemeSet) was already present in the codebase via OnceLock, no change needed.
…lone
Three optimizations targeting the hottest paths:
1. Static regex compilation + CompileId helpers (types.rs)
- Move RE_EVAL_WITH_KEY and RE_SEED_NSPID to module-level Lazy statics
- Add normalize_attempt() for None->Some(0) migration
- Add collapse_attempt() for unconditional attempt reset to 0
(used in compilation_metrics and metrics_index lookups)
2. Eliminate double JSON parse per log line (lib.rs) — HIGHEST IMPACT
- Parse each line as Envelope only once (was: Value + Envelope)
- Shortraw (raw.jsonl) output now built by parsing as Value separately,
inserting glog metadata, and re-serializing with sorted keys
- Substring-based key-conflict detection as early bail-out before parse
- Net effect: ~50% reduction in JSON parsing for the main loop
3. Avoid Vec<OutputFile> clone in CompilationMetrics (lib.rs, parsers.rs)
- Two-phase borrow pattern: immutable slice borrow for parse, then
mutable access for result processing
- Changed CompilationMetricsParser.output_files from &Vec to &[OutputFile]
- Eliminates clone of entire output file list per metrics entry
Output is byte-for-byte identical to baseline across all test logs.
Instead of loading the full input log into a String and passing it through
ParseOutput, the CLI now copies raw.log directly via std::fs::copy().
For a 500MB log, this saves ~500MB+ of heap allocation (String + UTF-8
validated copy). fs::copy uses kernel-level zero-copy (sendfile/copy_file_range).
Changes:
- lib.rs: Removed fs::read_to_string(path) and raw.log ParseOutput entry
- cli.rs: Added fs::copy(log_path, output_dir.join("raw.log")) after
writing all ParseOutput entries
Note: raw.log is not listed in the non-breaking contract as a guaranteed
ParseOutput entry. Library callers using parse_path() directly will no
longer find raw.log in the returned Vec and should copy the input file
themselves if needed.
Detect .gz extension on input files and transparently decompress using flate2::read::GzDecoder. This is purely additive — existing .log files work identically. Changes: - lib.rs: Wrap file reader in GzDecoder when path ends in .gz, using Box<dyn io::Read> for unified handling - cli.rs: Copy as raw.log.gz (not raw.log) for gzip inputs - cli.rs: Accept .log.gz files in --all-ranks-html rank log discovery (tries .log.gz suffix before .log) - Cargo.toml: Add flate2 = "1.0" dependency Tests: 3 new integration tests covering library-level gzip parsing, CLI raw.log.gz copying, and all-ranks .log.gz discovery. Verified: gzip output is byte-for-byte identical to uncompressed baseline for all test logs.
yushangdi
reviewed
Apr 2, 2026
Cargo.toml
Outdated
Contributor
There was a problem hiding this comment.
can you also update the version here?
Contributor
|
can you summarize the optimziations you made in this PR? |
Performance improvements in this release: - ~39% faster parsing (median) on large logs - ~32% less memory usage - Transparent gzip input support (.gz files) - fs::copy for raw.log (avoids loading entire file into memory)
Author
|
Updated the summary, each commit also has the description of changes within it. |
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
2 participants
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.
Performance and usability improvements for tlparse, yielding ~39% faster parsing and ~32% less memory on large logs.
Performance:
Features:
Infrastructure:
PR is split into smaller commits, each having appropriate description of changes.
Here are the benchmark results before vs after on a 327MB file.