mirror of
https://github.com/OPCFoundation/UA-.NETStandard.git
synced 2026-10-18 22:55:08 +08:00
romanett
28f1401324
Profiling the reference server under load showed most request-path CPU spent spinning on contended locks, plus several avoidable allocations per request or per node. This PR removes them, switches the reference server to Server GC, and adds the load harness used to measure it. ## Results Paired A/B against `master` (3 alternating rounds, medians, 20 sessions, workstation GC on both sides, 24-core machine): | Scenario | Throughput | p50 | p90 | Server CPU / op | Server alloc / op | Lock contention | |---|---|---|---|---|---|---| | Read, 1 value | **+45%** (94k → 137k req/s) | −20% | −46% | −37% | −39% | −94% | | Read, 100 values | **+40%** | −26% | −17% | −24% | −25% | +130% ¹ | | Read, 100 values, SignAndEncrypt | **+48%** | −36% | −20% | −30% | −26% | +106% ¹ | | Browse (Server object) | **+33%** | −31% | −12% | −13% | −14% | −40% | | Write, 100 values | +13% | −12% | −4% | −10% | −19% | −23% | | Data-change subscriptions | ±0 ² | ≈ | ≈ | −14% | −17% | −29% | ¹ The faster server spends more time in the remaining per-node locks. ² Throughput is bounded by the writer rate in this scenario. Separately, Server GC instead of workstation GC doubles Read-100 throughput (21k → 44k req/s) and cuts p90 from 2.1 ms to 0.6 ms, because workstation GC paused the server for about half the wall time under load. The working set grows from ~210 MB to ~430 MB. ## Changes - **`ObjectPool<T>`**: lock-free slot array. `ConcurrentBag.Count` took every per-thread lock on each `Return` of the pooled response value-task source, once per request. - **`FairRequestQueue`**: - drops redundant gate entries: the fail-fast pre-checks and the stopped check read volatile fields and are re-checked under the gate; the owner count is folded into selection; - coalesces wake pulses with a pending flag, so the bounded signal channel's lock is only taken when no token is outstanding. Wake-count semantics in the tests are unchanged. - **`DefaultServerResourceIsolationProvider`**: - compares peer addresses without `GetAddressBytes()` allocations; - builds the unsecured verification context only when a classifier is configured. - **TCP listener / channel**: caches the certificate DER copies per channel, keyed on the certificate instance. `X509Certificate2.RawData` returns a fresh copy on every access, and the listener read it for every request. - **`EndpointBase`**: resolves the request `ActivitySource` once instead of `Assembly.GetCallingAssembly()` per request. - **Node managers**: - one `ServerSystemContext` copy per request (cached on the `OperationContext`) for per-node callbacks: permission metadata, node metadata, view membership, method lookup; - removes a discarded copy in `GetReferenceDescriptionAsync`; - per-request operation caches are plain `Dictionary` instead of a `ConcurrentDictionary` with 24 lock objects; - the validation cache is built in one pass; - fixes zero-capacity error lists. - **`ResourceManager`**: translation lookups read an immutable snapshot that is rebuilt after writes. Browse serialised on the global lock for every display name. - **`MonitoredItem`**: shares the default `DataChangeFilter` instead of allocating one per sampled value. - **Reference server**: `ServerGarbageCollection=true`; new "Garbage collector mode" section in `docs/ServerScalability.md`. - **`perf/`**: `ServerLoadHarness` (out-of-process server plus load driver, server-side CPU/GC/alloc/contention deltas, optional `dotnet-trace` capture), `TraceAnalyzer` (TraceEvent-based CPU/allocation/contention aggregation), and paired A/B scripts. Not part of `UA.slnx`. ### Behaviour notes for custom node managers - The `IDictionary<NodeId, NodeState>` operation cache handed to overrides is now a `Dictionary`. It was a `ConcurrentDictionary`, but only one request flow uses it. - Within one request, per-node callbacks of a node manager share one `ServerSystemContext` copy. ### Later commits on this PR - **Permission checks without `NodeMetadata`** (`9068684c3`): the role-permission and access-restriction validators run on an internal `PermissionMetadata` struct; `AsyncCustomNodeManager` fills it on the stack from the same attributes and namespace defaults. Browse's reference filter reads 3 attributes per target instead of 13. Read 100 values: 120.6 → 95.7 KB/request, ~21k → 26k req/s; Browse 67.6 → 51.8 KB, 26.7k → 29.8k req/s. - **Read-only channel evidence** (`936b28d76`, `dc638ab9f`): `SecureChannelContext.ClientChannelCertificate`, `ServerChannelCertificate` and `ChannelThumbprint` are `ByteString`, so the cached per-channel certificate buffer is shared without a copy and cannot be altered by handlers or classifiers. **Public API change**, see the new section in `docs/MigrationGuide.md`. The isolation provider's revalidation snapshot still copies. - **Session lookup** (`936b28d76`): `SessionManager.TryGetSessionContext` takes the candidate snapshot without the bindings lock; the validating check under the lock is unchanged (2 instead of 4 global lock acquisitions per request). - Unit tests for `ObjectPool`, the per-request system context cache, the translation snapshot, `SecureChannelContext` null semantics and a stopped `FairRequestQueue`; `perf/` follows the repository conventions. ## Testing - `Opc.Ua.Server.Tests` (net10.0): 8386 passed before the `ByteString` change. After it, 13 secure-session and isolation tests failed; both causes are fixed in `dc638ab9f` and those tests (25 with their variants) pass. A full rerun of the Server, Sessions, Core, Redundancy, LDS and GDS suites on the final commit is in progress. - Full solution builds for net10.0 and net48; `dotnet format whitespace` clean. ### o6-automation opcua-benchmarks (sampling, deep pipelines) Replicated both tests of the [o6-automation benchmark](https://github.com/o6-automation/opcua-benchmarks) in the harness (`--scenario counters`, and `read` with 10 clients × N outstanding), paired against `origin/master`: | Test | master | this PR | |---|---|---| | Subscription sampling, 100 Hz, every item ≥ 99 % of samples, 1 item | FAIL (75–94 %) | PASS (100 %) | | same, 1024 items | FAIL (64–73 %) | PASS (99.9 %) | | same, capacity | — | between 4096 and 6144 items (~410–520k samples/s) | | Read, 10 clients × 100 outstanding, None | stalls (0–1.7k req/s) | 121–131k req/s | | Read, 10 clients × 100 outstanding, SignAndEncrypt | stalls (0 req/s) | 89–91k req/s | | Read, 10 clients × 1000 outstanding | channels faulted (`BadTcpNotEnoughResources`) | 0 errors, ~45–54k req/s | - **Fixed-rate sampling** (`30b769d13`): the sampling loop waited one interval after each sample, adding timer overshoot and sampling time to every cycle; samples now have absolute deadlines. - **Async response send** (`cdac5c5ca`): `SendResponse` blocked thread-pool workers on the channel gate; with many requests in flight the receive loop that was handed the gate waited for a pool thread and the channel stalled. The TCP, Kestrel TCP and WebSocket listeners now use `SendResponseAsync`. - **Backpressure instead of channel fault** (`191cd2ba7`): server channels stop reading while more than 100 responses are queued and resume at 50, instead of faulting the channel. Client channels keep reading. At 10,000 outstanding requests throughput is ~50k req/s with or without the limit engaged, so the remaining gap to the 100-per-client case is not the backpressure. ## Not changed - `SentMessageQueue` deep-clones every `NotificationMessage`. Retained messages return their pooled payloads on acknowledgement while the response copy may still be queued for send, and a client can acknowledge a sequence number before receiving it; the clone keeps a recycled payload from reaching another session's response. Removing it needs a send-completion hook from the transport. - Generated responses allocate a default `ResponseHeader` that is replaced (~2-3 % of per-request allocation). Fixing it means making structure fields of every generated type lazily initialized, a generator-wide change to default values, equality and encoding. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
3.7 KiB
3.7 KiB
Server performance tooling
Tools for measuring server latency, throughput and GC cost under load. They are not part of
the product build (not in UA.slnx) and target net10.0 only.
| Path | Purpose |
|---|---|
ServerLoadHarness/ |
Starts a reference server in a child process and drives load against it from the parent process. |
TraceAnalyzer/ |
Aggregates the dotnet-trace files the harness records (CPU, allocations, lock contention). |
ab.sh, ab-summary.ps1 |
Paired A/B runs of two harness builds, and the median summary. |
The server runs in its own process, so the server-side numbers (CPU, allocated bytes, GC count
and pause time, Monitor.LockContentionCount) exclude the load generator. The NUnit
LoadTest fixture in tests/Opc.Ua.Sessions.Tests checks correctness under load; this harness
measures it.
Build
CustomTestTarget=net10.0 dotnet build perf/ServerLoadHarness -c Release
dotnet build perf/TraceAnalyzer -c Release
dotnet tool install -g dotnet-trace
Run
perf/ServerLoadHarness/bin/Release/net10.0/ServerLoadHarness.exe run --scenario read --nodes 100 --sessions 20 --duration 15 --warmup 8
| Option | Default | Meaning |
|---|---|---|
--scenario |
read |
read, write, browse (Server object) or sub (data-change subscriptions plus a writer). |
--sessions |
20 | Concurrent sessions. |
--nodes |
100 | Values per Read/Write request, or monitored items per subscription. |
--inflight |
1 | Outstanding requests per session (closed loop). |
--security |
none |
none, sign or encrypt (Basic256Sha256). |
--subs, --pub, --write-interval |
5, 100, 100 | Subscriptions per session, publishing interval and writer interval in ms (sub). |
--duration, --warmup |
20, 10 | Measurement window and warm-up in seconds. |
--diag, --audit |
0 | Server diagnostics and auditing. |
--trace <file> |
Record the measurement window with dotnet-trace. |
|
--trace-profile |
cpu |
cpu (sampled thread time), alloc (allocation ticks), contention or gc. |
--out <file> |
Append the result block to a file. |
Environment variables prefixed SERVER_ are passed to the server process without the prefix,
for example SERVER_DOTNET_gcServer=1 to compare garbage collector modes.
The result block reports client throughput and latency percentiles, the server deltas for the
window, and the server's allocated bytes and CPU time per request (or per notification).
Allocated bytes per operation are deterministic and the most reliable signal. On a shared
machine, throughput and latency vary from run to run, so compare builds with ab.sh rather
than with two separate runs.
Analyse a trace
dotnet perf/TraceAnalyzer/bin/Release/net10.0/TraceAnalyzer.dll alloc trace.nettrace 40
dotnet perf/TraceAnalyzer/bin/Release/net10.0/TraceAnalyzer.dll cpu trace.nettrace 40 Monitor.Enter_Slowpath
dotnet perf/TraceAnalyzer/bin/Release/net10.0/TraceAnalyzer.dll contention trace.nettrace 15
The sampled thread-time profile also samples blocked threads. The cpu mode drops samples
whose leaf frame is a known wait (thread-pool semaphore, I/O completion poll, GC suspension),
so the remaining samples approximate on-CPU time. Spin-waiting in Monitor.Enter_Slowpath or
Lock.TryEnterSlow is real CPU and shows up there.
A/B comparison
Build the baseline harness in a second worktree, then alternate the two builds:
git worktree add --detach C:/uabase origin/master
cp -r perf C:/uabase/ && (cd C:/uabase && CustomTestTarget=net10.0 dotnet build perf/ServerLoadHarness -c Release)
cd perf && BASELINE=C:/uabase/perf/ServerLoadHarness/bin/Release/net10.0/ServerLoadHarness.exe ./ab.sh
pwsh -File ab-summary.ps1