[Japabu]

Files issues/many-osbench-rows-on-the-amd-bench-are-bimodal-per-boot.md: a measurement-infrastructure defect that #1129's A/Bs exposed.

Problem

On the AMD bench many osbench rows land in one of two modes per boot, base images included, mostly by placement:

• the pipe writer thread in the other LLC from its reader;
• a joined thread in the other LLC;
• spawned children in the other LLC from their parent;
• a client and its server split.

A decision rule that compares arms by three boots' p50 ranges therefore compares mode draws. ab9 and ab10 blocked #1129 and #1153 partly on such rows. The issue lists the rows, their measured causes (from per-round diaries and the kernel log's spawn and exit lines) and two that have no cause yet.

Resulting behaviour

None: issues only. It proposes, for the next measurement-infrastructure item:

1. Where a row's placement is a fact of the system, osbench reports it per sample or round (each side's CPU and LLC, a same|split class), so rules compare within a class.
2. Where it is not what the row means to measure, the row controls it, for example as two pinned rows.
3. Rows with no placement cause yet report the inputs that might decide them.

Evidence: ab10, ab9.

Checks

None run: one Markdown file.

🤖 Generated with Claude Code