ARM64 benchmarks and many-instance plan

Current ARM64 data

The homepage publishes ARM64 Wago-versus-wazero rows measured on one ARM64 machine. The structured copy is in project.json. Published instantiation values describe Go heap bytes or allocation objects created during instantiation; they are not total per-instance memory.

Instantiation allocation figures exclude guest linear memory, native code mappings, and native virtual-memory reservations.

Many-instance benchmark status

Status: protocol published; results not measured. Wago does not currently claim “80 instances under X MiB RSS.” Results will appear here only after the raw samples and reproduction command are committed.

Instance counts

N = 1, 8, 80, 120, with compile-once and fresh-process runs separated.

Workloads

• No-memory scalar module
• 128 KiB initial linear memory
• 1 MiB initialized data
• 16 MiB active working set
• Host-call-heavy and compute-heavy loops
• Create/close churn and long-lived warm instances

Required outputs

• Total throughput, per-instance fairness, and p50/p95/p99 latency
• RSS, PSS, and virtual memory reported separately
• Minor/major page faults, Go heap, GC CPU/time, and compile-once cost
• Incremental memory per live instance and memory returned after close
• A run under an explicit cgroup memory limit

Until those fields are present, this page deliberately contains no synthetic headline.