Evaluation Reports Index

This page summarizes the archived evaluation work that informed the maintained documentation set. The goal here is to retain the substantive findings in production documentation tone instead of pointing readers only to raw report paths.

Profiling Evaluations

The profiling reports established that Copper profiling is operationally useful at three levels:

  • aggregate summaries for high-level workload shape

  • top-path outputs for bounded hotspot analysis

  • full-path outputs for forensic environment and library-path diagnosis

The version4 profiling evaluation also confirmed that the pre-destroy snapshot path preserves the same essential signal as the final outputs, which is valuable when shutdown becomes noisy.

The maintained summaries now live in:

Registration and Scaling Evaluations

The registration-readiness reports showed that the original narrow race was real, but that the larger scaling bottleneck was provider-registration skew.

The maintained conclusions are:

  • local provider registration is not the same as remote parent readiness

  • the readiness handshake is correct and worth keeping

  • improved probing policy reduces visible race symptoms

  • provider-registration skew is the dominant remaining large-scale bottleneck

Those findings are now integrated into:

Metadata ENOENT TTL Evaluation

The -E evaluation established that:

  • very small TTL windows such as E100 are too small at 128 nodes

  • most of the gain arrives by roughly E1000 to E2000

  • E2000 is the strongest balanced default from the measured dataset

  • E10000 gave the best raw time in the 128-node test, but only narrowly

That evaluation should be read as an older debug-phase reference. It was collected before the later stable startup path was in place, so it is most useful as a rough estimate and tuning guide for the older code line rather than as the final production baseline.

Those findings are now integrated into:

How to Use the Maintained Pages

Use the maintained pages under docs/source for current operational guidance. The archived experiment directories under exp_results remain useful as source material, but the production documentation should be the first reading path for most users.

Sample Markdown Outputs

For full-file rendered examples of the generated Markdown artifacts, see: