Reduce overhead in converting resolutions #11660

konstin · 2025-02-20T11:02:43Z

Solving spent a chunk of its time just converting resolutions, the left two blocks:

These blocks are ResolverOutput::from_state with 1.3% and ForkState::into_resolution with 4.1% of resolver thread runtime for apache airflow universal.

We reduce the overhead spent in those functions, to now 1.1% and 2.1% of resolver time spend in those functions by:

Commit 1: Replace the hash set for the edges with a vec in ForkState::into_resolution. We deduplicate edges anyway when collecting them, and the hash-and-insert was slow.

Commit 2: Reduce the distribution clonign in ResolverOutput::from_state by using an Arc.

The same profile excerpt for the resolver with the branch (note that there is now an unrelated block between the two we optimized):

Wall times are noisy, but the profiles show those changes as improvements.

$ hyperfine --warmup 2 "./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal" "./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal"
Benchmark 1: ./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal
  Time (mean ± σ):      99.1 ms ±   3.8 ms    [User: 111.8 ms, System: 115.5 ms]
  Range (min … max):    93.6 ms … 110.4 ms    29 runs
 
Benchmark 2: ./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal
  Time (mean ± σ):      97.1 ms ±   4.3 ms    [User: 114.8 ms, System: 112.0 ms]
  Range (min … max):    90.9 ms … 112.4 ms    29 runs
 
Summary
  ./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal ran
    1.02 ± 0.06 times faster than ./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal

konstin · 2025-02-20T11:03:14Z

crates/uv-resolver/src/resolver/mod.rs

-    pub(crate) edges: FxHashSet<ResolutionDependencyEdge>,
+    pub(crate) edges: Vec<ResolutionDependencyEdge>,


That's the big perf improvement

Building the `HashSet` was a block in the profile (2.6% of the resolver thread). We don't need to hash at all since we're deduplicating edges when reading them anyway. The benchmark isn't too clear but the profile show it as an improvement. ``` Benchmark 1: ./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal Time (mean ± σ): 102.0 ms ± 6.8 ms [User: 116.4 ms, System: 114.6 ms] Range (min … max): 93.1 ms … 115.8 ms 32 runs Benchmark 2: ./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal Time (mean ± σ): 98.6 ms ± 4.4 ms [User: 114.0 ms, System: 112.7 ms] Range (min … max): 93.3 ms … 112.4 ms 31 runs Summary ./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal ran 1.03 ± 0.08 times faster than ./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal ```

konstin · 2025-02-20T11:04:38Z

crates/uv-installer/src/plan.rs

@@ -393,7 +394,7 @@ pub struct Plan {

    /// The distributions that are not already installed in the current environment, and are
    /// not available in the local cache.
-    pub remote: Vec<Dist>,
+    pub remote: Vec<Arc<Dist>>,


This is isn't necessary but we avoid the clone as we already have Arcs here.

In `ResolverOutput::from_state`, we spend a lot of time in cloning `ResolvedDist`. We can reduce that time by using an `Arc` around the `Dist`s in `ResolvedDist`.

This MR contains the following updates: | Package | Update | Change | |---|---|---| | [astral-sh/uv](https://github.com/astral-sh/uv) | patch | `0.6.0` -> `0.6.3` | MR created with the help of [el-capitano/tools/renovate-bot](https://gitlab.com/el-capitano/tools/renovate-bot). **Proposed changes to behavior should be submitted there as MRs.** --- ### Release Notes <details> <summary>astral-sh/uv (astral-sh/uv)</summary> ### [`v0.6.3`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#063) [Compare Source](astral-sh/uv@0.6.2...0.6.3) ##### Enhancements - Allow quotes around command-line options in `requirement.txt files` ([#11644](astral-sh/uv#11644)) - Initialize PEP 723 script in `uv lock --script` ([#11717](astral-sh/uv#11717)) ##### Configuration - Accept multiple `.env` files in `UV_ENV_FILE` ([#11665](astral-sh/uv#11665)) ##### Performance - Reduce overhead in converting resolutions ([#11660](astral-sh/uv#11660)) - Use `SmallString` on `Hashes` ([#11756](astral-sh/uv#11756)) - Use a `Box` for `Yanked` on `File` ([#11755](astral-sh/uv#11755)) - Use a `SmallString` for the `Yanked` enum ([#11715](astral-sh/uv#11715)) - Use boxed slices for hash vector ([#11714](astral-sh/uv#11714)) - Use install concurrency for bytecode compilation too ([#11615](astral-sh/uv#11615)) ##### Bug fixes - Avoid installing duplicate dependencies across conflicting groups ([#11653](astral-sh/uv#11653)) - Check subdirectory existence after cache heal ([#11719](astral-sh/uv#11719)) - Include uppercase platforms for Windows wheels ([#11681](astral-sh/uv#11681)) - Respect existing PEP 723 script settings in `uv add` ([#11716](astral-sh/uv#11716)) - Reuse refined interpreter to create tool environment ([#11680](astral-sh/uv#11680)) - Skip removed directories during bytecode compilation ([#11633](astral-sh/uv#11633)) - Support conflict markers in `uv export` ([#11643](astral-sh/uv#11643)) - Treat lockfile as outdated if (empty) extras are added ([#11702](astral-sh/uv#11702)) - Display path separators as backslashes on Windows ([#11667](astral-sh/uv#11667)) - Display the built file name instead of the canonicalized name in `uv build` ([#11593](astral-sh/uv#11593)) - Fix message when there are no buildable packages ([#11722](astral-sh/uv#11722)) - Re-allow HTTP schemes for Git dependencies ([#11687](astral-sh/uv#11687)) ##### Documentation - Add anchor links to arguments and options in the CLI reference ([#11754](astral-sh/uv#11754)) - Add link to environment marker specification ([#11748](astral-sh/uv#11748)) - Fix missing a closing bracket in the `cache-keys` setting ([#11669](astral-sh/uv#11669)) - Remove the last edited date from documentation pages ([#11753](astral-sh/uv#11753)) - Fix readme typo ([#11742](astral-sh/uv#11742)) ### [`v0.6.2`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#062) [Compare Source](astral-sh/uv@0.6.1...0.6.2) ##### Enhancements - Add support for constraining build dependencies with `tool.uv.build-constraint-dependencies` ([#11585](astral-sh/uv#11585)) - Sort dependency group keys when adding new group ([#11591](astral-sh/uv#11591)) ##### Performance - Use an `Arc` for index URLs ([#11586](astral-sh/uv#11586)) ##### Bug fixes - Allow use of x86-64 Python on ARM Windows ([#11625](astral-sh/uv#11625)) - Fix an issue where conflict markers could instigate a very large lock file ([#11293](astral-sh/uv#11293)) - Fix duplicate packages with multiple conflicting extras declared ([#11513](astral-sh/uv#11513)) - Respect color settings for log messages ([#11604](astral-sh/uv#11604)) - Eagerly reject unsupported Git schemes ([#11514](astral-sh/uv#11514)) ##### Documentation - Add documentation for specifying Python versions in tool commands ([#11598](astral-sh/uv#11598)) ### [`v0.6.1`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#061) [Compare Source](astral-sh/uv@0.6.0...0.6.1) ##### Enhancements - Allow users to mark platforms as "required" for wheel coverage ([#10067](astral-sh/uv#10067)) - Warn for builds in non-build and workspace root pyproject.toml ([#11394](astral-sh/uv#11394)) ##### Bug fixes - Add `--all` to `uvx --reinstall` message ([#11535](astral-sh/uv#11535)) - Fallback to `GET` on HTTP 400 when attempting to use range requests for wheel download ([#11539](astral-sh/uv#11539)) - Prefer local variants in preference selection ([#11546](astral-sh/uv#11546)) - Respect verbatim executable name in `uvx` ([#11524](astral-sh/uv#11524)) ##### Documentation - Add documentation for required environments ([#11542](astral-sh/uv#11542)) - Note that `main.py` used to be `hello.py` ([#11519](astral-sh/uv#11519)) </details> --- ### Configuration 📅 **Schedule**: Branch creation - At any time (no schedule defined), Automerge - At any time (no schedule defined). 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever MR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this MR and you won't be reminded about this update again. --- - [ ] If you want to rebase/retry this MR, check this box --- This MR has been generated by [Renovate Bot](https://github.com/renovatebot/renovate).

Solving spent a chunk of its time just converting resolutions, the left two blocks: ![image](https://github.com/user-attachments/assets/6f266440-c6e2-447c-ad7f-f92244f9d09b) These blocks are `ResolverOutput::from_state` with 1.3% and `ForkState::into_resolution` with 4.1% of resolver thread runtime for apache airflow universal. We reduce the overhead spent in those functions, to now 1.1% and 2.1% of resolver time spend in those functions by: Commit 1: Replace the hash set for the edges with a vec in `ForkState::into_resolution`. We deduplicate edges anyway when collecting them, and the hash-and-insert was slow. Commit 2: Reduce the distribution clonign in `ResolverOutput::from_state` by using an `Arc`. The same profile excerpt for the resolver with the branch (note that there is now an unrelated block between the two we optimized): ![image](https://github.com/user-attachments/assets/e36c205d-2cf8-4fe6-a2dd-3020c0515922) Wall times are noisy, but the profiles show those changes as improvements. ``` $ hyperfine --warmup 2 "./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal" "./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal" Benchmark 1: ./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal Time (mean ± σ): 99.1 ms ± 3.8 ms [User: 111.8 ms, System: 115.5 ms] Range (min … max): 93.6 ms … 110.4 ms 29 runs Benchmark 2: ./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal Time (mean ± σ): 97.1 ms ± 4.3 ms [User: 114.8 ms, System: 112.0 ms] Range (min … max): 90.9 ms … 112.4 ms 29 runs Summary ./uv-branch pip compile --no-progress scripts/requirements/airflow.in --universal ran 1.02 ± 0.06 times faster than ./uv-main pip compile --no-progress scripts/requirements/airflow.in --universal ```

konstin commented Feb 20, 2025

View reviewed changes

konstin force-pushed the konsti/reduce-edge-hashing branch from be3ea7e to 1d7b6f9 Compare February 20, 2025 11:03

konstin commented Feb 20, 2025

View reviewed changes

Optimize ResolvedDist clones with Arc

f465616

In `ResolverOutput::from_state`, we spend a lot of time in cloning `ResolvedDist`. We can reduce that time by using an `Arc` around the `Dist`s in `ResolvedDist`.

konstin force-pushed the konsti/reduce-edge-hashing branch from 1d7b6f9 to f465616 Compare February 20, 2025 11:21

zanieb added the performance Potential performance improvement label Feb 20, 2025

zanieb approved these changes Feb 20, 2025

View reviewed changes

charliermarsh merged commit ae916cf into main Feb 20, 2025
74 checks passed

charliermarsh deleted the konsti/reduce-edge-hashing branch February 20, 2025 20:13

BrewTestBot mentioned this pull request Feb 24, 2025

uv 0.6.3 Homebrew/homebrew-core#208817

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reduce overhead in converting resolutions #11660

Reduce overhead in converting resolutions #11660

konstin commented Feb 20, 2025

konstin Feb 20, 2025

konstin Feb 20, 2025

		pub(crate) edges: FxHashSet<ResolutionDependencyEdge>,
		pub(crate) edges: Vec<ResolutionDependencyEdge>,

Reduce overhead in converting resolutions #11660

Reduce overhead in converting resolutions #11660

Conversation

konstin commented Feb 20, 2025

konstin Feb 20, 2025

Choose a reason for hiding this comment

konstin Feb 20, 2025

Choose a reason for hiding this comment