Correlating GHC Cost-Center Profiles with ThreadScope Eventlogs for Unified Bottleneck Analysis

Integration Challenge

GHC provides two complementary profiling mechanisms: cost-center profiling (-prof -fprof-auto) for source-level attribution of time and allocation, and the runtime eventlog (-eventlog) for ThreadScope visualization of thread scheduling, garbage collection, and spark behavior. Both can be enabled simultaneously during compilation, yet they emit independent data streams with no shared correlation identifiers.

Constraints and Uncertainty

Cost-center profiling introduces 2–10× overhead and can distort lazy evaluation, while eventlog files grow to hundreds of megabytes for long runs. The GHC community has discussed extending the eventlog format to embed cost-center identifiers, but proposals face backward-compatibility and eventlog-size concerns. No stable toolchain natively merges the .prof report with the .eventlog timeline; developers manually align hotspots across separate runs, risking misattribution when lazy forcing sites differ from definition sites.

Assuming GHC 9.6+ with -threaded and multiple capabilities (+RTS -N), what is the current feasibility of a single-run workflow that binds cost-center entries to eventlog timestamps?

1. Does any released GHC version or third-party tool emit a combined profile where each eventlog record carries a cost-center stack identifier?
2. If not, what minimal runtime changes would be required to add cost-center tags to the eventlog without prohibitive overhead?
3. Are there documented workarounds that reliably correlate the two data sources post-hoc for parallel Haskell programs?