Skip to the content.

Cache-budget concentration allocation alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s concentration-weighted allocator, equal-share allocation, and request-volume proportional allocation. LMCache, Mooncake, vLLM, and SGLang retain zero measurements until real shared cache pools execute the common workload; #6159 tracks those witnesses.

Same-workload contract

Every arm allocates 1,200 cache bytes at top-K one across three tenants: a concentrated 9:1 workload, a flat 1:1 workload, and an unmeasured workload. The oracle requires exact budget conservation, concentration preference over the flat tenant, and a protected 400-byte flat share for the unmeasured tenant. Equal-share ignores concentration; volume proportional allocation can starve the unmeasured tenant.

Complete engine runs replay an equivalent request trace and report captured cache value, budget conservation, starved tenants, allocation and request latency, throughput, cache bytes, CPU/RSS/network/storage, and total cost.

Arm Class Local availability Honest result
fak native concentration-weighted allocation native yes conserves 1,200 bytes and protects unmeasured share
equal-share cache allocation tuned baseline yes conservative but ignores concentration
request-volume proportional allocation tuned baseline yes demand-aware but can starve unmeasured tenant
fak + LMCache first-class integration no real shared cache runtime required
fak + Mooncake first-class integration no real shared cache runtime required
vLLM cache-aware routing external no zero measurements
SGLang HiCache and cache-aware scheduling external no zero measurements

Adapters and allocator-shaped mocks are not cache-engine witnesses.

Local native witness

go test ./internal/vcachecal -bench BenchmarkAllocateByConcentration -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 439.0, 472.8, 471.5, 534.3, 567.7 ns/op. Median: 472.8 ns/three-bucket allocation, 712 B/op, 10 allocs/op. This is allocator overhead, not cache-engine request latency.

Reproduce

go test ./internal/vcachecal -run TestCompareLocalKeepsAllocationAlternativesExplicit
go test ./internal/vcachecal -bench BenchmarkAllocateByConcentration -benchmem -run '^$' -count=5
go test ./internal/nativebench