Memory optimization handoff plan

English summary

This page summarizes the handoff plan. The detailed evidence, file locations, risks, and acceptance criteria are in the full Traditional Chinese plan.

Changes already merged

PR #4 includes:

• Bounded channel pools for Aster's 16–128 KiB allocator slabs, limiting retained pool entries after connection bursts.
• Compact DNS cache entries for eligible A/AAAA-only replies using netip.Addr values. Replies requiring full DNS message semantics retain the full-message representation.

These changes do not by themselves establish a measured reduction in whole-process resident memory. Aster's allocator and the dependency's sing buffer pool are separate.

September 7 validation update

The performance report now includes Windows/Linux verification and seven-round WSL2 A/B results. Median RSS with 4,096 DNS cache entries fell 2.5%, with overlapping ranges; 1,000 TCP connections did not use less RAM. Large-pool Get/Put and full DNS-message cache hits became slower; full-message hit allocations rose from 252 B / 5 allocs to 516 B / 10 allocs. Natural-GC TCP heap profiles were insufficient to complete P0-1 attribution. The planned items below remain unaccepted.

Priorities

1. Attribute idle TCP memory with heap profiles for SOCKS-to-DIRECT and SOCKS-to-TLS-proxy scenarios before changing relay buffers. Preserve splice, read-waiter, protocol headroom, counters, and connection lifecycle behavior.
2. Investigate right-sizing small queued UDP packets instead of retaining a full receive slab for each packet.
3. Measure QUIC receive-window pressure before considering smaller defaults for with_low_memory; investigate connection-panel snapshot allocations.
4. Measure LRU overhead, geodata startup GC costs, and tracker metadata retention before choosing optimizations.
5. Investigate slow log subscribers as a robustness issue; observe provider-loading memory peaks without changing behavior first.

These are planned investigations, not completed optimizations.

Measurement and safety rules

• Use fresh processes and seven interleaved runs for before/after comparisons. Report medians and ranges.
• Measure Windows working set/private bytes or Linux RSS/high-water mark separately; do not treat them as interchangeable or extrapolate RSS from benchmark B/op.
• Collect heap profiles to establish allocation ownership. Include relevant benchmarks and correctness tests.
• Do not tune GOGC or GOMEMLIMIT, add a periodic memory reclaimer, or force GC to improve reported numbers. Without explicit user approval, retain the current no-GC-intervention policy.
• Preserve cache capacities, TTLs, stale DNS behavior, fake-IP reverse lookup, protocol semantics, and public JSON fields.
• Keep each optimization in a separate PR, with evidence and regression results.

Verification checklist

• Build normal and with_low_memory variants.
• Run the full Go test suite and race tests for changed packages.
• Run formatting checks, go vet, and golangci-lint.
• Compare relevant benchmarks with seven runs and -cpu=1.
• Record whole-process measurements and heap attribution before claiming memory savings.

See the performance reference for existing measurements and the full plan for per-item acceptance criteria.