[kevinherron]

Part 3 of 6 (#2118 to #2123) of a stack of server performance changes, found by profiling the Milo server in the o6-automation/opcua-benchmarks suite.

Problem

Every service that takes a list of operations routes each one to the address space fragment that owns it, then puts the results back in request order. For a Read of 100 Variables that is 100 lookups plus the access check's internal Read of 400 more attributes. Both steps ran a Stream pipeline per call:

• AddressSpaceComposite.getAddressSpace ran a filter().findFirst() over the fragments for every operation, and CompositeAddressSpaceFilter ran an anyMatch over its children for every nested test.
• GroupMapCollate.groupMapCollate, used by 33 call sites across the services, ran groupingBy and three collect calls per request.

In a profile of 10 clients reading batches of 100 Variables, taken with part 2 (#2119) applied, the routing lookups were about 12% of server CPU and GroupMapCollate's stream collection about 11%, more than the attribute reads themselves. Separately, ManagedAddressSpace.read and write built their DEBUG log arguments for every operation even at INFO: getNodeId() through the attribute filter chain, and an AttributeId lookup mapped through an Optional. That was about 2% of server CPU.

Change

Two commits.

Dispatch operations to address space fragments without Streams. getAddressSpace and every CompositeAddressSpaceFilter method use plain loops. GroupMapCollate groups item indices in a LinkedHashMap and collates the results into an array. Routing is unchanged: the first fragment whose filter accepts an operation handles it, and an empty fragment is built only when none does (the fix from #2098 is kept).

GroupMapCollate keeps its contract, which had no tests: each mapper is called once with a new mutable list of its group's items in request order; results come back in request order; a mapper that returns the wrong number of results is rejected, even if it changed the list it was given; a null key still throws NullPointerException. One thing changes: groups are now visited in first-seen order rather than HashMap order, which was never specified.

Build ManagedAddressSpace read and write log arguments only for DEBUG. The DEBUG messages are unchanged. With DEBUG off, a custom attribute filter no longer sees the incidental getNodeId() call those arguments made.

Results

Same benchmark setup as the rest of the stack: open62541 C clients, default configuration, Ryzen 9 7950X with the server on 16 CPUs, median of 6 runs.

Workload	1.2	Before this PR	With this PR	Change
async, 10 clients, batch of 100 read	57,564	60,708	118,137	+95%
sync, 1 client, batch of 100 read	6,408	6,495	9,890	+52%
async, 10 clients, scalar read	540,578	543,592	595,458	+10%
sync, 1 client, scalar read	34,382	32,268	36,604	+13%

Tests

• New GroupMapCollateTest (sdk-core): results follow request order across interleaved groups, including a group larger than the initial index capacity; each mapper sees its group once, in order; a mapper that clears its input still has its results collated; a mapper that returns too few results is rejected; a null key is rejected.
• AddressSpaceCompositeRoutingTest covers first-match routing and the lazily built empty fragment.

Verification

• GroupMapCollateTest: 6 tests pass. AddressSpaceCompositeRoutingTest: 8 tests pass.
• Full build with tests on the top of the stack: 5,792 tests, no failures, 4 skipped.
• Reviewed independently by GPT-6.1-Sol (Codex) in two rounds. It checked all synchronous GroupMapCollate callers for dependence on group order, list type, or key semantics. Its round-1 finding, that a mapper which mutated its input could defeat the result-count check, is fixed and covered by the two mutation tests.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] This PR appears safe to merge; no actionable issues were found.

Summary

This PR replaces stream-based address-space routing and grouped result collection with loops.

• Grouped request results keep their original order with less bookkeeping.
• Address-space routing checks fragments with direct loops.
• Read and write logs skip detail work when debug is off.

No actionable issues were found. Tests were inspected, not run.

_{Reviews (1) · Last reviewed commit: "Build ManagedAddressSpace read and write..." · Reviewed by Greptile}