build-parallelism — скилл для Claude Code: разбор параллельной сборки

---
name: build-parallelism
description: "Analyze an MSBuild solution, solution filter, or Build.proj that schedules multiple project files. USE FOR: measured idle worker nodes, `-m` that does not scale, a serial ProjectReference critical path, graph builds, or an MSBuild task that builds a Projects list. Requires at least two distinct project files whose scheduling or throughput must be analyzed. DO NOT USE for non-MSBuild build systems."
license: MIT
---

## Diagnose a slow parallel build (start here)

Work this checklist in order — it targets the usual root cause (a serial
dependency chain that no number of cores can parallelize):

1. **Check how many worker nodes actually built the solution.** Capture a
   binlog with `dotnet build /bl:{}` (PowerShell: `dotnet build '-bl:{}'`) and
   inspect the node timeline to see whether more than one node ran. Every
   `dotnet` CLI command built on the SDK's MSBuild-forwarding path (`build`,
   `msbuild`, `test`, `pack`, `publish`, etc.) enables `/maxcpucount` (multiple
   nodes) by default; `-m:1` forces a single node. A direct `MSBuild.exe`
   invocation is sequential by default, so pass `-m` explicitly to enable
   parallel worker nodes there.
2. **Find the critical path.** From the binlog, read per-project timings and the
   node timeline. If total build time ≈ the sum of the projects on one
   dependency chain, that chain — not CPU count — is the bottleneck.
3. **Name the chain explicitly**, e.g. `Core → Api → Web → Tests`. A long serial
   chain stays serial no matter how large `-m` is, because each project waits on
   its predecessor.
4. **Look for unnecessary `ProjectReference` edges** that lengthen the chain — a
   reference that only needs build order (not the output assembly), or one that
   could be a `PackageReference`, forces serialization it doesn't need.
5. **Recommend flattening**: break false dependencies so independent projects
   build concurrently, and consider `/graph` for better scheduling.

## MSBuild Parallelism Model

- `/maxcpucount` (or `-m`): number of worker nodes (processes)
- Every `dotnet` CLI command (`build`, `msbuild`, `test`, `pack`, `publish`,
  etc.) passes `/maxcpucount` by default via the SDK's shared MSBuild-forwarding
  path; `-m:1` overrides it with one node
- A direct `MSBuild.exe` invocation is sequential by default and requires an
  explicit `-m`
- Recommended: `-m` without a number = use all logical processors
- Each node builds one project at a time
- Projects are scheduled based on dependency graph

## Project Dependency Graph

- MSBuild builds projects in dependency order (topological sort)
- Critical path: longest chain of dependent projects determines minimum build time
- Bottleneck: if project A depends on B, C, D and B takes 60s while C and D take 5s, B is the bottleneck
- Diagnosis: replay binlog to diagnostic log with `performancesummary` and check Project Performance Summary — shows per-project time; grep for `node.*assigned` to check scheduling
- Wide graphs (many independent projects) parallelize well; deep graphs (long chains) don't

## Graph Build Mode (`/graph`)

- `dotnet build /graph` or `msbuild /graph`
- What it changes: MSBuild constructs the full project dependency graph BEFORE building
- Benefits: better scheduling, avoids redundant evaluations, enables isolated builds
- Limitations: all projects must use `<ProjectReference>` (no programmatic MSBuild task references)
- When to use: large solutions with many projects, CI builds
- When NOT to use: projects that dynamically discover references at build time

## Optimizing Project References

- Reduce a `<ProjectReference>` only after proving it is unnecessary; a slow but
  valid dependency must stay in the graph
- Use `<ProjectReference ... SkipGetTargetFrameworkProperties="true">` to avoid extra evaluations
- `<ProjectReference ... ReferenceOutputAssembly="false">` for build-order-only dependencies
- Consider if a ProjectReference should be a PackageReference instead (pre-built NuGet)
- Use `solution filters` (`.slnf`) to build subsets of the solution

## BuildInParallel

- `<MSBuild Projects="@(ProjectsToBuild)" BuildInParallel="true" />` in custom targets
- Without `BuildInParallel="true"`, MSBuild task batches projects sequentially
- Ensure `/maxcpucount` > 1 for this to have effect

## Multi-threaded MSBuild Tasks

- Individual tasks can run multi-threaded within a single project build
- Tasks implementing `IMultiThreadableTask` can run on multiple threads
- Tasks must declare thread-safety via `[MSBuildMultiThreadableTask]`

## Analyzing Parallelism with Binlog

### Primary: binlog MCP (preferred)

Use the **binlog MCP server** (`Microsoft.AITools.BinlogMcp`, exposed under the `binlog` MCP namespace):

1. Use expensive_projects tool → find the slowest projects and compare individual vs total build time
2. Use expensive_targets tool → find bottleneck targets
3. Use project_target_times tool → drill into a specific project's target-level timing
4. Ideal: build time should be much less than sum of project times (parallelism)
5. If build time ≈ sum of project times: too many serial dependencies, or one slow project blocking others

### Fallback: text-log replay (when MCP is unavailable)

Step-by-step:

1. Replay the binlog: `dotnet msbuild build.binlog -noconlog -fl "-flp:v=diag;logfile=full.log;performancesummary"`
2. Check Project Performance Summary at the end of `full.log`
3. Ideal: build time should be much less than sum of project times (parallelism)
4. If build time ≈ sum of project times: too many serial dependencies, or one slow project blocking others
5. `grep 'Target Performance Summary' -A 30 full.log` → find the bottleneck targets
6. Consider splitting large projects or optimizing the critical path

## CI/CD Parallelism Tips

- Use `-m` in CI (many CI runners have multiple cores)
- Consider splitting solution into build stages for extreme parallelism
- Use build caching (NuGet lock files, deterministic builds) to avoid rebuilding unchanged projects
- `dotnet build /graph` works well with structured CI pipelines