Table of Contents

JNet: performance

This page reports benchmark results for the core JNet interop primitives: JVM method invocation from .NET and JVM→.NET callback latency. All benchmarks run on GitHub Actions runners and are repeated automatically on each release across supported .NET and JDK versions.

Results are reported for three JCOBridge versions — 2.6.6, 2.6.7+, and 2.6.9 — across two runtime combinations. See JCOBridge release notes for details.

Note

Benchmarks are run on shared GitHub-hosted runners. Absolute numbers reflect that environment and should be read comparatively rather than as absolute throughput figures for a dedicated host.

For the full BenchmarkDotNet results across all supported .NET versions, JDK vendors, and platforms see:


Test environments

Three environments are used across this page:

Parameter x86-64 EPYC (stopwatch) x86-64 (BenchmarkDotNet) ARM64 (BenchmarkDotNet)
Runner GitHub Actions ubuntu-22.04, AMD EPYC 9V45 96-Core GitHub Actions ubuntu-24.04, AMD EPYC 7763 2-Core GitHub Actions ubuntu-24.04, Neoverse-N2 4-Core
Measurement Stopwatch, 1 000 000 iterations BenchmarkDotNet, statistically rigorous BenchmarkDotNet, statistically rigorous
.NET versions .NET 8 / .NET 10 .NET 8 / .NET 10 .NET 8 / .NET 10
JDK versions Temurin 17 / Temurin 25 Multiple vendors — see Latest benchmark results Multiple vendors — see Latest benchmark results
Note

The 2.6.6 and 2.6.7+ baselines were collected on earlier ubuntu-latest runners and are preserved as historical reference. The x86-64 EPYC stopwatch results (2.6.9 section) use the AMD EPYC 9V45 dedicated runner and represent the lowest latency observed on x86-64. The BenchmarkDotNet results use shared ubuntu-24.04 runners on both platforms and are statistically rigorous; see the live dashboard for up-to-date numbers.


What is measured

JVM object creation

Measures the cost of creating a JVM object from .NET via three constructor resolution strategies:

  • NewEmpty — resolves the constructor by argument type matching.
  • NewWithSignature — resolves the constructor by JNI signature string.
  • DeclaredNewEmpty — uses a pre-declared constructor reference.

Field access

Measures the cost of getting and setting JVM instance and static fields from .NET. Primitive fields (e.g. int) are transferred as-is; reference fields (e.g. String) require marshalling between JVM and CLR representations.

JVM method invocation from .NET

Measures the round-trip latency of calling a JVM method from .NET through JNet, with two resolution strategies and two method signatures.

Resolution strategies:

  • Invoke — the method is identified by .NET-side type matching against input arguments on every call. The JVM descriptor is cached after first resolution, but argument type validation is re-evaluated on the .NET side at each invocation.
  • InvokeWithSignature (IWS) — the method is identified by name and JNI signature string. Argument validation is delegated to the JVM, eliminating the .NET-side type matching cost.

Method signatures (feedback):

  • feedback = false — method takes no arguments and returns void. Measures pure invocation overhead.
  • feedback = true — method takes a boolean argument and returns the same boolean. Measures the additional cost of argument passing and return value marshalling across the JNI boundary.

Both static and instance method variants are tested.

Array and varargs invocation

Measures the cost of passing arrays or varargs to JVM methods from .NET, across three element types and three sizes (10, 1 000, 100 000 elements):

  • InvokeIntArrayFixed — fixed-length int[] array.
  • InvokeVarArgsWholeArrayint[] passed as a varargs array (single JNI call).
  • InvokeStringArrayFixed — fixed-length String[] array — each element requires individual JVM↔CLR string marshalling.
  • InvokeVarArgsSpreadElements — each element passed as an individual varargs argument — worst case for element-by-element overhead.
  • InvokeIntParam — single int parameter, used as baseline.

Multi-parameter invocation

Measures the overhead of passing multiple mixed-type arguments in a single JVM method call from .NET.

Callback: TestPredicateRoundTrip (.NET → JVM → .NET)

A .NET-initiated test: .NET triggers a JVM call which immediately fires a callback back into .NET. Provides a controlled start-time marker and measures the full bidirectional round-trip. In real usage the JVM initiates the event — see TestPredicateSustained for the realistic reference.

Callback: TestPredicateSustained (JVM → .NET, sustained)

A JVM-initiated test: .NET sends a single start command to JVM, which then fires 1 000 000 callback events toward the CLR autonomously without returning control to .NET. After all events are fired, the JVM returns and .NET measures the total elapsed time. Aside from the single startup call, this measures the pure cost of receiving a sustained stream of JVM-originated events — the scenario matching real-world usage (e.g. Kafka Streams functional interfaces, AWT event listeners).

Both callback tests share two configuration axes (2.6.7+ only):

byIndex — event trigger identification:

  • byIndex = false — the event is identified on the CLR side by a string key lookup.
  • byIndex = true — the event is identified on the CLR side by a numeric index. In both cases, JVM object arguments are retrieved as JVM objects after the trigger is received.

Two-level early-discard filter (ListenerShallManageEvent, 2.6.7+):

JCOBridge 2.6.7+ introduces two overloads of ListenerShallManageEvent on the JNet callback base class, forming a two-gate filter applied before full event handling. Both gates receive the event as a numeric index — no string conversion is performed unless explicitly requested via the name-based delegate variants.

First gate — bool ListenerShallManageEvent(int eventIndex): called before any argument data is read from the JVM. The return value:

  • false (continueFirstCheck = false) — discard immediately: no data is read, the handler is not invoked.
  • true (continueFirstCheck = true) — proceed to the second gate.

The first gate is driven by one of the following, evaluated in order:

  • ListenerShallManageEventIndex (Func<int, bool>) — fastest: receives the raw event index, no string conversion.
  • ListenerShallManageEventName (Func<string, bool>) — receives the event name, resolved via ConvertListenerEventIndexToEventName.
  • Override of ListenerShallManageEvent(int) — virtual, for subclass-based filtering.
  • Default: returns true (all events proceed).

Second gate — bool ListenerShallManageEvent(int eventIndex, object data): called after raw argument data is available but before full event processing and handler dispatch. Allows lightweight inspection of the raw payload without paying the cost of full conversion. The return value:

  • false (continueSecondCheck = false) — discard after raw-data inspection: the registered handler is not invoked.
  • true (continueSecondCheck = true) — proceed normally: full argument conversion and handler invocation.

The second gate is driven by one of the following, evaluated in order:

  • ListenerShallManageEventIndexWithData (Func<int, object, bool>) — receives the raw event index and raw data.
  • ListenerShallManageEventNameWithData (Func<string, object, bool>) — receives the event name and raw data.
  • Override of ListenerShallManageEvent(int, object) — virtual.
  • Default: returns true.

The combination continueFirstCheck = false, continueSecondCheck = true is never reached and is not tested. Default for both gates is true (full processing).


JVM object creation (BenchmarkDotNet, ubuntu-24.04, Temurin 17 / Temurin 25)

Method x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
NewEmpty 1.169 µs 1.109 µs 1.046 µs 1.046 µs
NewWithSignature 1.173 µs 1.108 µs 1.033 µs 1.033 µs
DeclaredNewEmpty 1.158 µs 1.121 µs 1.033 µs 1.033 µs

All three strategies are equivalent in cost — the JNI boundary crossing dominates over the resolution overhead. No allocation occurs on the .NET side. ARM64 constructors are ~5-10% faster than x86-64 on the same JDK version.


Field access (BenchmarkDotNet, ubuntu-24.04, Temurin 17 / Temurin 25)

Method x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
GetInstanceIntField 100.4 ns 76.2 ns 100.4 ns 100.4 ns
SetInstanceIntField 120.5 ns 81.4 ns 121.7 ns 121.7 ns
GetInstanceStringField 104.5 ns 73.6 ns 113.8 ns 113.8 ns
SetInstanceStringField 631.7 ns 605.9 ns 674.2 ns 674.2 ns
GetStaticIntField 122.7 ns 78.2 ns 114.3 ns 114.3 ns
SetStaticIntField 127.7 ns 79.8 ns 110.7 ns 110.7 ns
GetStaticStringField 101.2 ns 70.3 ns 97.7 ns 97.7 ns
SetStaticStringField 638.3 ns 625.2 ns 667.3 ns 667.3 ns
GetInstanceIntFieldGeneric 137.1 ns 82.2 ns 105.9 ns 105.9 ns

Primitive field access is symmetric (~75–140 ns get/set across platforms). String field writes are ~6–8× more expensive than reads — setting a String field requires converting the .NET string to a JVM String object, an allocation on the JVM heap plus a full string copy. Avoid repeated writes to JVM String fields in hot paths; prefer int/long/boolean fields where performance matters.


Method invocation (BenchmarkDotNet, ubuntu-24.04, Temurin 17 / Temurin 25)

For full multi-vendor comparison see Latest benchmark results.

Method x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
InvokeStaticEmpty 273.4 ns 275.7 ns 296.0 ns 298.6 ns
InvokeStaticEmptyWithSignature 251.4 ns 267.7 ns 291.9 ns 305.7 ns
InvokeStaticWithFeedback 517.6 ns 568.0 ns 549.6 ns 530.6 ns
InvokeInstanceEmpty 278.8 ns 283.2 ns 300.0 ns 310.0 ns
InvokeInstanceWithFeedback 533.5 ns 587.4 ns 534.0 ns 547.2 ns

On x86-64 with .NET 8 / T17, InvokeWithSignature is ~8% faster than Invoke for static methods; the gap narrows on .NET 10. On ARM64 both strategies are within 5% of each other. The advantage of InvokeWithSignature is more pronounced under load and with complex argument types — see the live dashboard for a full multi-vendor breakdown.


Array and varargs invocation (BenchmarkDotNet, ubuntu-24.04, Temurin 17 / Temurin 25)

Method Elements x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
InvokeIntParam (baseline) 559 ns 595 ns 538 ns 539 ns
InvokeIntArrayFixed 10 1,301 ns 1,188 ns 1,183 ns 1,131 ns
InvokeIntArrayFixed 1 000 1,981 ns 1,999 ns 1,900 ns 1,859 ns
InvokeIntArrayFixed 100 000 62,896 ns 67,119 ns 60,633 ns 60,633 ns
InvokeVarArgsWholeArray 100 000 63,741 ns 67,268 ns 60,484 ns 60,483 ns
InvokeStringArrayFixed 10 6,412 ns 6,262 ns 7,204 ns 7,204 ns
InvokeStringArrayFixed 1 000 483,498 ns 462,731 ns 551,710 ns 551,710 ns
InvokeStringArrayFixed 100 000 59,122,284 ns 53,696,166 ns 57,603,003 ns 57,603,003 ns
InvokeVarArgsSpreadElements 100 000 102,160,469 ns 95,098,483 ns 80,829,867 ns 80,829,867 ns

int[] scales linearly with size at ~0.6–0.7 ns per element after the fixed JNI boundary cost. String[] costs ~0.5–0.75 µs per element regardless of size — each element requires an individual JVM→CLR Unicode conversion and allocation. At 100 000 elements that accumulates to ~54–59 ms vs ~60–67 µs for integers.

Note

The scaling difference between int[] and String[] reflects the fundamental cost of string marshalling: each String element requires an individual JVM→CLR Unicode conversion and allocation, while each int element is a direct memory copy. When passing large collections of string data across the boundary, consider encoding them as a single binary payload (byte[] or JCOBridgeStream<byte>) and parsing on the receiving side, or restructuring the API to avoid per-element crossings.


Preallocated JVM object invocation (BenchmarkDotNet, ubuntu-24.04, Temurin 17 / Temurin 25)

When JNet wrapper objects (Java.Lang.String, or any class generated by JNetReflector) are created in advance and reused across calls, passing them to a JVM method requires only transferring the native object reference — no marshalling, no allocation at call time. These benchmarks measure that path.

Method Elements x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
InvokeSingleStringPreallocated 1 577 ns 303 ns 589 ns 583 ns
InvokeStringArrayPreallocated 10 2,011 ns 1,167 ns 1,949 ns 1,958 ns
InvokeStringArrayPreallocated 1 000 60,540 ns 31,922 ns 59,528 ns 61,262 ns
InvokeStringArrayPreallocated 100 000 7,454,725 ns 5,217,584 ns 9,241,561 ns 7,947,215 ns

Comparison at 100 000 elements (x86-64 .NET 10 / T25):

Approach 100 000 elements Per-element cost
InvokeIntArrayFixed 67,119 ns ~0.6 ns
InvokeStringArrayPreallocated (preallocated JNet objects) 5,217,584 ns ~52 ns
InvokeStringArrayFixed (.NET string, marshalled) 53,696,166 ns ~537 ns

Key observations:

  • Single preallocated object — ~300-590 ns, essentially the same cost as a no-argument method invocation. No marshalling occurs: only the native JVM object pointer is passed through the boundary.
  • Preallocated array — eliminates string marshalling (~8-10× faster than passing .NET string[] at 100 000 elements) but still pays a per-element cost for the type analysis layer: for each array element the engine runs a type dispatch chain to identify the object kind (IJVMBridgeBaseInstance), extracts the native JVM pointer, and builds the corresponding factory. No marshalling occurs, but the per-element .NET-side dispatch is not free (~52-92 ns per element depending on platform and runtime).
  • vs int[] — even with preallocated objects, passing 100 000 object references costs ~80-140× more than passing 100 000 integers, because integers are transferred as a contiguous memory block while each object element must go through the type dispatch chain.
  • Conclusion — preallocating JNet wrapper objects is the right approach when the same JVM objects are used repeatedly across many calls (e.g. a fixed set of configuration objects, enum constants, or frequently reused strings). For truly bulk data transfer, prefer JCOBridgeStream<T> with primitive types.

Method x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
InvokeMultiParam 1.403 µs 1.390 µs 1.330 µs 1.330 µs

Multi-parameter invocation costs roughly 4–5× a no-argument call, reflecting the overhead of boxing and passing each additional argument across the JNI boundary. Keep argument lists short for frequently called methods.


In 2.6.6, the ListenerShallManageEvent filter and the native byIndex trigger mechanism are not yet available.

Static method invocation

Resolution feedback .NET 8 / T17 .NET 10 / T25
Invoke false 0.661 µs 0.602 µs
IWS false 0.494 µs 0.414 µs
Invoke true 0.901 µs 0.803 µs
IWS true 0.686 µs 0.522 µs

Instance method invocation

Resolution feedback .NET 8 / T17 .NET 10 / T25
Invoke false 0.579 µs 0.490 µs
IWS false 0.468 µs 0.379 µs
Invoke true 0.856 µs 0.764 µs
IWS true 0.638 µs 0.535 µs

Adding a boolean argument and return value (feedback = true) adds ~45–55% overhead with Invoke and ~35–40% with IWS, reflecting JNI argument marshalling cost.

Callback

Test byIndex .NET 8 / T17 .NET 10 / T25
RoundTrip false 6.945 µs 6.338 µs
Sustained false 6.116 µs 5.548 µs

Sustained is the realistic reference for JVM-originated callback cost: ~6.1 µs (.NET 8 / T17) and ~5.5 µs (.NET 10 / T25).


JCOBridge 2.6.7+

JCOBridge 2.6.7+ introduces the two-level ListenerShallManageEvent filter and the native byIndex trigger mechanism. General interop improvements reduce baseline overhead across all test types.

Note

byIndex = true is still simulated on the JVM side by invoking a dedicated class method rather than the interface @Override. The CLR-side numeric index resolution is fully active; a JVM dispatch difference (class method vs interface method) remains. The byIndex = false rows use the real interface override and are directly comparable between versions.

Static method invocation

Resolution feedback .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
Invoke false 0.517 µs −22% 0.480 µs −20%
IWS false 0.356 µs −28% 0.335 µs −19%
Invoke true 0.609 µs −32% 0.575 µs −28%
IWS true 0.435 µs −37% 0.419 µs −20%

Instance method invocation

Resolution feedback .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
Invoke false 0.349 µs −40% 0.301 µs −39%
IWS false 0.295 µs −37% 0.274 µs −28%
Invoke true 0.552 µs −36% 0.511 µs −33%
IWS true 0.448 µs −30% 0.452 µs −15%

Callback: TestPredicateRoundTrip

byIndex continueFirstCheck continueSecondCheck .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
false false false 1.106 µs 1.040 µs
true ¹ false false 0.452 µs 0.435 µs
false true false 1.126 µs 1.074 µs
true ¹ true false 0.502 µs 0.456 µs
false true true 5.794 µs −16% 5.318 µs −16%
true ¹ true true 5.023 µs −28% 4.628 µs −27%

¹ byIndex = true simulated on the JVM side — see note above.

Callback: TestPredicateSustained

byIndex continueFirstCheck continueSecondCheck .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
false false false 0.601 µs −90% 0.468 µs −92%
true ¹ false false 0.045 µs 0.041 µs
false true false 0.625 µs −90% 0.493 µs −91%
true ¹ true false 0.074 µs 0.067 µs
false true true 5.098 µs −17% 4.725 µs −15%
true ¹ true true 4.467 µs −27% 4.141 µs −25%

¹ byIndex = true simulated on the JVM side — see note above.

The realistic JVM-originated callback baseline (full processing, byIndex = false) is 5.1 µs (.NET 8 / T17) and 4.7 µs (.NET 10 / T25).


JCOBridge 2.6.9

JCOBridge 2.6.9 delivers further improvements across all test types through interop layer optimizations, lazy initialization of event data (TypedEventData), and a pooled BatchState buffer that eliminates per-event array allocation in the dispose fast-scope path. Results collected on ubuntu-22.04 / AMD EPYC 9V45 96-Core.

Note

byIndex = true is still simulated on the JVM side — see note in the 2.6.7+ section above.

Static method invocation

Resolution feedback .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
Invoke false 0.313 µs −53% 0.265 µs −56%
IWS false 0.248 µs −50% 0.250 µs −40%
Invoke true 0.471 µs −48% 0.482 µs −40%
IWS true 0.311 µs −55% 0.305 µs −42%

Instance method invocation

Resolution feedback .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
Invoke false 0.189 µs −67% 0.176 µs −64%
IWS false 0.162 µs −65% 0.178 µs −53%
Invoke true 0.389 µs −55% 0.390 µs −49%
IWS true 0.287 µs −55% 0.307 µs −43%

Callback: TestPredicateRoundTrip

byIndex continueFirstCheck continueSecondCheck .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
false false false 0.709 µs −37% 0.768 µs −21%
true ¹ false false 0.240 µs 0.252 µs
false true false 1.028 µs −6% 0.935 µs −4%
true ¹ true false 0.401 µs 0.451 µs
false true true 3.157 µs −55% 3.278 µs −48%
true ¹ true true 2.607 µs 2.713 µs

¹ byIndex = true simulated on the JVM side — see note above.

Callback: TestPredicateSustained

byIndex continueFirstCheck continueSecondCheck .NET 8 / T17 vs 2.6.6 .NET 10 / T25 vs 2.6.6
false false false 0.379 µs −94% 0.366 µs −93%
true ¹ false false 0.035 µs 0.030 µs
false true false 0.505 µs −92% 0.522 µs −91%
true ¹ true false 0.144 µs 0.152 µs
false true true 2.674 µs −56% 2.750 µs −50%
true ¹ true true 2.324 µs 2.278 µs

¹ byIndex = true simulated on the JVM side — see note above.

The realistic JVM-originated callback baseline (full processing, byIndex = false) reaches 2.7 µs on both .NET 8 and .NET 10 — a −56% reduction over 2.6.6.

Callback: BenchmarkDotNet comparison (ubuntu-24.04, Temurin 17 / Temurin 25)

TestPredicateSustained results. For full multi-vendor breakdown see Latest benchmark results.

byIndex continueFirstCheck continueSecondCheck x86-64 .NET 8 / T17 x86-64 .NET 10 / T25 ARM64 .NET 8 / T17 ARM64 .NET 10 / T25
false false false 511.5 ns 497.2 ns 574.8 ns 584.7 ns
true ¹ false false 40.0 ns 30.5 ns 44.2 ns 42.6 ns
false true false 689.2 ns 692.8 ns 781.4 ns 800.4 ns
true ¹ true false 185.6 ns 204.4 ns 262.9 ns 255.7 ns
false true true 3,917.6 ns 4,081.8 ns 4,059.6 ns 4,514.7 ns
true ¹ true true 3,263.7 ns 3,133.3 ns 3,611.1 ns 3,627.0 ns

¹ byIndex = true simulated on the JVM side — see note in the 2.6.7+ section.

The x86-64 first-gate (byIndex = true, F,F) reaches 30.5 ns on .NET 10 / T25 — identical to the x86-64 EPYC stopwatch result (~30–35 ns) and confirming the measurement across two independent methodologies. ARM64 first-gate is 42–44 ns. Full processing on both platforms converges at ~3.1–4.5 µs (BenchmarkDotNet shared runner), with the EPYC stopwatch showing lower numbers (~2.3–2.7 µs) reflecting dedicated hardware advantage.

The three distinct operating points:

First gate only (continueFirstCheck = false) — event discarded before any data is read:

  • byIndex = false: ~0.38–0.37 µs — string key lookup.
  • byIndex = true: ~35 ns (.NET 8) / ~30 ns (.NET 10)ListenerShallManageEventIndex path, pure integer check, no string conversion. ~28–33 M events/sec.

Second gate (continueFirstCheck = true, continueSecondCheck = false) — raw data available for inspection, handler not invoked:

  • byIndex = false: ~0.51–0.52 µs.
  • byIndex = true: ~144 ns (.NET 8) / ~152 ns (.NET 10) — includes raw data retrieval from JVM.

Full processing (continueFirstCheck = true, continueSecondCheck = true): ~2.7 µs (byIndex = false), ~2.3 µs (byIndex = true).


Bulk data transfer at the JVM↔CLR boundary

JCOBridge 2.6.9 introduces JCOBridgeDirectBuffer<T> (wrapping a JVM DirectByteBuffer) and JCOBridgeStream<T> (wrapping a JVM native array), both with T : unmanaged. Both types expose ToStream() (backed by UnmanagedMemoryStream), ReadOnlySpan<T>, and .NET Framework-compatible shims.

Note

Tests run in a single process without isolation. Memory is pre-allocated once per size step; 100 iterations measure only access/transfer cost. Future benchmarks will use DotNetBenchmark with process isolation for statistically rigorous results.


Array transfer — JCOBridgeStream<T>

A JVM byte[] of the given size is pre-allocated once per size step. Each iteration retrieves the data via three APIs:

  • Invoke<byte[]> — allocates a .NET byte[] and copies JVM array data into it. Behavior is the same in both standard and HPA editions; the underlying transfer path can be switched to a lower-overhead mode for small arrays (see below).
  • AreEqualChunked — reads via JCOBridgeStream<byte> in 4096-byte chunks without allocating a full copy.
  • AsSpan — obtains a ReadOnlySpan<byte> from JCOBridgeStream<byte>:
    • Standard edition: performs an internal copy regardless of forceRawMemory (forceRawMemory is a no-op in standard)
    • HPA, forceRawMemory=false: accesses JVM array memory directly — no copy
    • HPA, forceRawMemory=true: accesses JVM array memory with GC pinned for the duration — no copy, lowest latency for large arrays; use with care (no JVM allocations or blocking operations during the pinned window)

Standard edition — AsSpan latency (µs, 100 iterations, default path / optimized path)

The optimized transfer path reduces per-call overhead for small arrays dramatically. For large arrays (≥1 MB) the bottleneck shifts to memory bandwidth and the optimized path offers no advantage.

Size .NET 8 default .NET 8 optimized .NET 10 default .NET 10 optimized
10 B 6.3 1.4 6.8 1.5
1 KB 3.5 1.4 2.5 1.1
10 KB 4.3 2.6 3.6 1.6
100 KB 77.2 59.6 79.6 62.6
1 MB 289.2 212.8 248.7 192.2
10 MB 2,152.6 2,213.4 1,449.3 1,501.9
100 MB 35,962.8 59,541.3 14,958.8 20,502.4

The optimized path is beneficial only for small arrays (≤100 KB). For large arrays it can be slower due to GC interaction — use the default path at scale.

HPA edition — AsSpan latency (µs, 100 iterations, before optimized path)

Size Standard HPA forceRaw=false HPA forceRaw=true vs Standard
10 B .NET 8 6.3 5.1 4.7 −25%
100 KB .NET 8 77.2 10.0 8.3 −89%
1 MB .NET 8 289.2 84.7 46.8 −84%
10 MB .NET 8 2,152.6 764.2 270.8 −87%
100 MB .NET 8 35,962.8 17,830.0 6,019.2 −83%
100 KB .NET 10 79.6 12.6 9.2 −88%
1 MB .NET 10 248.7 134.6 74.4 −70%
10 MB .NET 10 1,449.3 858.2 314.6 −78%
100 MB .NET 10 14,958.8 15,826.7 5,626.6 −62%

Key observations:

  • forceRawMemory=false eliminates the copy for ≥100 KB: −87/−88% at 100 KB, −84/−70% at 1 MB vs standard. For small arrays (≤10 KB) the difference is minor — overhead is dominated by per-call cost, not data movement.
  • forceRawMemory=true goes further: −83/−62% at 100 MB vs standard — the fastest path for large arrays. The additional gain over forceRawMemory=false is largest at 10 MB+ where the pinned access eliminates all intermediate buffering.
  • For small arrays (≤10 KB), all HPA paths are similar to standard — per-call overhead dominates over data transfer cost. Use the optimized transfer path (see standard table above) for small sizes.

ByteBuffer transfer — JCOBridgeDirectBuffer<T>

The benchmark reflects real-world usage: each iteration calls getByteBuffer() on the JVM side, which copies the pre-allocated array into a DirectByteBuffer before returning it to .NET. This includes both the JVM-side copy cost (heap → native memory) and the .NET-side read cost.

Note

JCOBridgeDirectBuffer<T> and EnableCritical / forceRawMemory have no effect on ByteBuffer access — a DirectByteBuffer already lives in native memory and is always accessed via direct pointer in both standard and HPA editions. Standard and HPA produce identical results.

Size ToArray .NET 8 AsSpan .NET 8 ToArray .NET 10 AsSpan .NET 10
10 B 198.5 µs 9.8 µs 192.5 µs 8.9 µs
100 B 8.4 µs 4.7 µs 8.4 µs 4.2 µs
1 KB 6.5 µs 4.4 µs 6.1 µs 4.5 µs
10 KB 13.9 µs 5.3 µs 15.3 µs 5.7 µs
100 KB 93.4 µs 12.3 µs 90.6 µs 13.5 µs
1 MB 552.0 µs 88.9 µs 398.0 µs 103.8 µs
10 MB 2,674.0 µs 881.8 µs 2,535.6 µs 815.8 µs
100 MB 44,096.1 µs 13,474.7 µs 20,339.0 µs 11,742.0 µs
  • AsSpan is consistently fastest: reads the native memory pointer directly with no .NET allocation.
  • ToArray at 10 B is expensive (198 µs) because getByteBuffer() has per-call JVM overhead that dominates at small sizes — getArray() + Invoke<byte[]> is faster for small payloads.
  • ToStream → Naive (full intermediate MemoryStream copy) is the slowest for large sizes — avoid above a few KB.
Note

ByteBuffer vs HPA array access: the AsSpan ByteBuffer cost at 100 MB (~13.5 ms on .NET 8) reflects two components: the JVM-side copy from heap to native memory (~7.5 ms) and the .NET read from native memory (~6 ms). The .NET read portion matches JCOBridgeStream AsSpan with HPA forceRawMemory=true (~6 ms), confirming that both operations ultimately read from native memory at the same speed. The ByteBuffer pattern adds the JVM-side copy overhead that HPA array access avoids entirely.


Summary

Test .NET 8 / T17 .NET 10 / T25
2.6.6 2.6.7+ 2.6.9 2.6.6 2.6.7+ 2.6.9
Static Invoke fb=false 0.661 0.517 (−22%) 0.313 (−53%) 0.602 0.480 (−20%) 0.265 (−56%)
Static IWS fb=false 0.494 0.356 (−28%) 0.248 (−50%) 0.414 0.335 (−19%) 0.250 (−40%)
Static Invoke fb=true 0.901 0.609 (−32%) 0.471 (−48%) 0.803 0.575 (−28%) 0.482 (−40%)
Static IWS fb=true 0.686 0.435 (−37%) 0.311 (−55%) 0.522 0.419 (−20%) 0.305 (−42%)
Instance Invoke fb=false 0.579 0.349 (−40%) 0.189 (−67%) 0.490 0.301 (−39%) 0.176 (−64%)
Instance IWS fb=false 0.468 0.295 (−37%) 0.162 (−65%) 0.379 0.274 (−28%) 0.178 (−53%)
Instance Invoke fb=true 0.856 0.552 (−36%) 0.389 (−55%) 0.764 0.511 (−33%) 0.390 (−49%)
Instance IWS fb=true 0.638 0.448 (−30%) 0.287 (−55%) 0.535 0.452 (−15%) 0.307 (−43%)
Sustained: full, byIndex=false 6.116 5.098 (−17%) 2.674 (−56%) 5.548 4.725 (−15%) 2.750 (−50%)
Sustained: full, byIndex=true ¹ 4.467 2.324 4.141 2.278
Sustained: 1st gate, byIndex=false 0.601 0.379 0.468 0.366
Sustained: 1st gate, byIndex=true ¹ 0.045 0.035 0.041 0.030
Sustained: 2nd gate, byIndex=false 0.625 0.505 0.493 0.522
Sustained: 2nd gate, byIndex=true ¹ 0.074 0.144 0.067 0.152

All values in µs. Percentages vs 2.6.6 where available. ¹ byIndex = true simulated on the JVM side.

Comparison with raw JNI overhead

The 30–35 ns figure for byIndex = true, first-gate discard (Sustained, 2.6.9) is worth contextualizing against published raw JNI benchmarks on dedicated hardware. Independent JMH benchmarks measure an empty JNI call at ~57 ns via JavaCPP and ~22 ns on a modern laptop for a minimal no-op native method (java-native-benchmark, Komanov 2022).

JNet's first-gate discard path (ListenerShallManageEventIndex) involves a JVM→CLR crossing, a numeric index lookup, and an immediate return — all on shared CI infrastructure. Reaching 30–35 ns per event places JNet within the range of raw JNI call overhead measured on dedicated bare-metal hardware, despite the additional CLR interop layer.


Guidance

  • Prefer InvokeWithSignature (IWS) over Invoke in hot paths on x86 — it avoids .NET-side type matching and consistently delivers 20–55% lower latency with arguments involved. On ARM64 the two strategies are nearly equivalent for no-argument methods; the advantage of IWS varies by JDK vendor (see dashboard).
  • Constructor resolution strategy does not matter — all three approaches (NewEmpty, NewWithSignature, DeclaredNewEmpty) cost ~1 µs. Choose whichever is most readable.
  • Primitive field access is cheap (~100–120 ns); string field writes are expensive (~670 ns, ~6–7× the getter cost). Avoid repeated JVM String field writes in hot paths — prefer int/long/boolean fields or aggregate string data into a single call.
  • Never pass large String[] arrays in hot paths — at 100 000 elements a String[] costs ~57 ms vs ~60 µs for an equivalent int[] (1000× difference). Prefer primitive arrays, or use JCOBridgeStream<T> for binary bulk transfer.
  • Varargs with spread elements are the worst case for per-element overhead — each argument is individually boxed. Pass arrays as arrays, not as spread varargs, when the element count is large.
  • Multi-parameter invocation costs ~1.3 µs — roughly 4× a no-argument call. Keep argument lists short for frequently called methods.
  • The realistic JVM-originated callback reference is Sustained, full processing, byIndex = false: ~2.7 µs on x86 EPYC (2.6.9), ~4.1–4.5 µs on ARM64 Neoverse-N2.
  • Use the two-level ListenerShallManageEvent filter for high-event-rate sources:
    • First gate (ListenerShallManageEventIndex) — discard before any data read: ~30–35 ns (x86), ~42–44 ns (ARM64).
    • Second gate (ListenerShallManageEventIndexWithData) — inspect raw data before handler: ~144–152 ns (x86), ~256–263 ns (ARM64).
  • For bulk data from JVM arrays, prefer JCOBridgeStream<T> with HPA forceRawMemory=true for large payloads — up to 83% faster than standard at 100 MB. For small arrays (≤10 KB), the optimized transfer path reduces per-call overhead to ~1–2 µs.
  • For DirectByteBuffer access, use AsSpan — zero-copy from native memory in all editions, ~16 GB/s at 100 MB. Avoid ToStream → Naive above a few KB.
  • The byIndex = true mechanism will deliver its full benefit on the full-processing path once the JVM-side simulation is replaced with real interface dispatch.
  • If your application runs callbacks at sustained high frequency, consider the JCOBridge HPA edition — it addresses GC-boundary instability under sustained JVM↔CLR call pressure and enables true zero-copy bulk array access.