[TianDi101]

Summary

Removes UMBP code that nothing calls, a test that has never been built, and five copies of the same helper. Most of it lost its last caller in the backend-agnostic refactor (#540). No behavior change. Net −1,370 lines across 45 files.

Each removal was checked three ways:

• Grep: a comment-stripped token count across every tracked file (C++, pybind, Python, tests, CMake, CI).
• Relocations: an objdump -r scan over every object file of a full UMBP build, to catch dead functions whose names are too common for grep.
• Delegation: a scan for virtual methods whose only call site is a same-named delegating override (e.g. GetLocationId, SetColdRead).

Dead files

File	Why it is dead
local/tiers/device_copy_run.h	Run planner "shared only by DRAMTier and its unit test"; both were removed in #540
local/tiers/device_gather.h, local/tiers/host_registration.h	"Backward-compatible" shims; nothing includes them (the real headers live in include/umbp/common/)
tests/cpp/umbp/local/test_spdk_proxy.cpp	Not in any CMakeLists.txt, at any point in history, so it has never been compiled

Dead functions and state

• PoolClient::ExecuteLocalGet (and GetAttemptOutcome) and PoolClient::CopyRangesToContiguous.
• SpdkSsdTier::Batch{Write,Read}Dma{Direct,Streaming}, about 490 lines.
• TierBackend::ReadPtr, BatchWriteRanges, GetLocationId, SetColdRead, plus their SSDTier, ShardedSsdTier and SpdkProxyTier overrides. Because nothing ever set it, SpdkProxyTier::cold_read_ was always false; the field and its branch are removed.
• SSDTier::last_io_status_ (with LastIoStatus/RememberStatus): written at 13 sites but never read. The writes ran concurrently from const read paths, so this also removes a latent data race.
• SPDK proxy and environment code:
  • SpdkEnv: the reactor accessors, PollIo, CleanupThreadLocalCtx, CleanupOnSpdkThread and bdev_.
  • ProxyShmRegion::ProbeExisting and its file-local HeartbeatStaleMs.
  • SpdkProxyTier::WaitForProxy and ProxyPollInterval. UMBP_SPDK_PROXY_POLL_INTERVAL_MS is dropped from runtime-env-vars.md because nothing reads it now.
  • TenantRegistry::FindTenant.
• Unused accessors, constants and fields:
  • HostTierRegistration::Covers and RegisteredBytes.
  • MasterClient::SupportsLogicalTiers, MoriIoEngine::LocalEngineKey, UMBPConfig::ResolveRole, kAutoRankId, kLocFieldMarker, LocalLocation, SegmentIndex::ClearKnownSegments, ThreadDoesBackgroundCopies.
  • Two "legacy alias" SPDK proxy names.
  • Write-only PeerSsdManager eviction counters (eviction is counted by the decorator and the pool) and write-only io_uring ring-entry fields.
• route_get_strategy.cpp: commented-out debug logging, plus SummarizeLocations, which only that logging used.

Duplicated code

• Five identical TierType ↔ ::umbp::TierType switch converters (three standalone/*.cpp files, peer_service.cpp, batch_resolve_codec.h) are now one inline pair in umbp/distributed/tier_proto.h. The static_cast versions in master_client.cpp behave differently on unknown values, so they are left alone.
• Three identical MatchedHashCount() bodies now call one shared CountDistinctHashes() in types.h. IUMBPClient::ExternalKvMatch and MasterClient::ExternalKvNodeMatch stay separate types because pybind registers them as two different Python classes.

Also

• .github/workflows/ci.yml: removes test_prefix_aware_eviction from the UMBP ctest regex. No test by that name has existed since refactor(umbp): make distributed mode backend- and transport-agnostic #540.

Kept on purpose

• TierBackend::ReadBatchRangesIntoPtr and the ranged_read capability: only tests call them today, but they are the building block for the still-open D1 phase of design-ssd-ranged-io.md.
• APIs used only by tests as inspection helpers (IMasterMetadataStore::LookupBlock, GetExternalKvCount, ...), the vendored SPDK offset_allocator, and the bench and run scripts (all referenced).
• Near-duplicates that have drifted apart, such as bench_pool_client_batch_{get,put}.cpp, or whose env parsing differs slightly. Merging those would be a refactor, not a deletion.

Test plan

Before/after comparison on n06-21 (local/mori-dev:sglang-0.5.19-*, gfx950). Build flags: BUILD_UMBP=ON, USE_REDIS_BACKEND=ON, BUILD_TESTS=ON, pybind mori_umbp_pybinds.

• Build: 290/290 targets for both main and this branch. Compiler warnings are identical (same 100).
• C++ unit tests (tests/cpp/umbp, 57 ctests): per-test outcomes are identical between main and this branch, 46 pass. The other 11 are PoolClient/IO-engine tests that need a GPU; they fail identically on both with rsmi_num_monitor_devices reported 0 GPUs, because the container had no GPU while the node's GPUs were fully busy with another job. CI runs them with a GPU.
• Python unit tests (pytest tests/python/umbp/): 105/105 on both, loading the freshly built libmori_umbp_pybinds.so (md5 verified). matched_hash_count() returns the same value on both match types.
• clang-format 20.1.8 and the repo pre-commit hooks: clean.
• SPDK-only files (spdk_env.cpp, spdk_ssd_tier.cpp, spdk_proxy_daemon.cpp): not compiled, because SPDK isn't available in the build image. Every edit there is a pure deletion of unreferenced code (braces balanced, no remaining references), but it still needs an SPDK build to confirm.
• GDS (hipfile): not compiled, but untouched.

🤖 Generated with Claude Code