Benchmark prs by jprendes · Pull Request #1529 · hyperlight-dev/hyperlight

jprendes · 2026-06-12T08:50:45Z

No description provided.

hyperlight-gh-bot · 2026-06-18T11:41:47Z

Benchmark Results

kvm / amd (Linux) (❌ *1.81x slower* → 🚀 **2.41x faster**)

function_call_serialization

	`serialize_function_call`	`deserialize_function_call`
	`7.91 ms` (❌ 1.81x slower)	`10.23 ms` (❌ 1.19x slower)

guest_calls

	`call_with_host_function`	`call_with_restore`	`call`	`different_thread`	`interrupt_latency`
`medium`	`23.92 µs` (✅ 1.07x slower)	`70.81 µs` (✅ 1.13x faster)	`12.77 µs` (✅ 1.02x slower)
`small`	`20.73 µs` (✅ 1.11x faster)	`65.39 µs` (✅ 1.08x faster)	`11.50 µs` (✅ 1.08x faster)
`default`	`23.58 µs` (✅ 1.02x slower)	`62.36 µs` (✅ 1.00x slower)	`12.91 µs` (✅ 1.02x slower)
				`9.82 ms` (❌ 1.48x slower)	`22.93 µs` (❌ 1.13x slower)
`large`	`24.52 µs` (✅ 1.07x slower)	`89.47 µs` (✅ 1.17x faster)	`12.97 µs` (✅ 1.02x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`596.90 ms` (✅ 1.07x slower)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`22.33 µs` (✅ 1.68x faster)	`21.02 µs` (🚀 1.86x faster)

sandboxes

	`create_uninitialized`	`create_initialized`	`create_uninitialized_and_drop`	`create_initialized_and_drop`
`large`	`28.23 ms` (🚀 2.31x faster)	`27.82 ms` (🚀 2.41x faster)	`29.26 ms` (🚀 2.25x faster)	`39.64 ms` (🚀 1.98x faster)
`medium`	`8.62 ms` (🚀 2.05x faster)	`9.84 ms` (🚀 1.97x faster)	`9.12 ms` (🚀 1.97x faster)	`24.22 ms` (✅ 1.27x faster)
`small`	`2.18 ms` (✅ 1.73x faster)	`5.11 ms` (✅ 1.03x slower)	`2.43 ms` (✅ 1.47x faster)	`18.61 ms` (❌ 1.19x slower)
`default`	`403.76 µs` (✅ 1.27x faster)	`2.59 ms` (❌ 1.43x slower)	`449.76 µs` (✅ 1.25x faster)	`18.50 ms` (❌ 1.59x slower)

shared_memory

	`copy_from_slice`	`fill`	`copy_to_slice`
`64MB`	`4.72 ms` (✅ 1.11x slower)	`3.03 ms` (❌ 1.20x slower)	`4.48 ms` (✅ 1.10x slower)
`1MB`	`40.44 µs` (✅ 1.02x slower)	`28.08 µs` (✅ 1.30x faster)	`38.55 µs` (✅ 1.07x slower)

snapshots

	`restore`	`create`
`medium`	`15.56 µs` (✅ 1.03x faster)	`28.75 ms` (✅ 1.21x faster)
`large`	`36.68 µs` (✅ 1.41x faster)	`100.35 ms` (✅ 1.40x faster)
`small`	`10.73 µs` (❌ 1.13x slower)	`2.93 ms` (✅ 1.20x faster)
`default`	`9.83 µs` (✅ 1.03x faster)	`302.77 µs` (✅ 1.28x faster)

Summary

Biggest gain: sandboxes/create_initialized/large — 🚀 2.41x faster
Worst regression: function_call_serialization/serialize_function_call — ❌ 1.81x slower

kvm / intel (Linux) (❌ *1.88x slower* → 🚀 **7.91x faster**)

function_call_serialization

	`serialize_function_call`	`deserialize_function_call`
	`7.01 ms` (❌ 1.23x slower)	`10.82 ms` (✅ 1.15x faster)

guest_calls

	`call_with_host_function`	`call_with_restore`	`call`	`different_thread`	`interrupt_latency`
`medium`	`39.32 µs` (✅ 1.04x slower)	`43.21 µs` (✅ 1.03x slower)	`20.59 µs` (✅ 1.07x slower)
`small`	`39.81 µs` (✅ 1.07x slower)	`37.25 µs` (✅ 1.01x faster)	`20.58 µs` (✅ 1.10x slower)
`default`	`36.34 µs` (✅ 1.03x faster)	`36.97 µs` (✅ 1.02x faster)	`21.09 µs` (✅ 1.09x slower)
				`10.90 ms` (❌ 1.64x slower)	`14.60 µs` (🚀 1.85x faster)
`large`	`37.04 µs` (✅ 1.05x slower)	`80.05 µs` (✅ 1.04x slower)	`19.39 µs` (✅ 1.00x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`653.71 ms` (✅ 1.04x faster)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`34.89 µs` (🚀 3.07x faster)	`34.42 µs` (🚀 3.24x faster)

sandboxes

	`create_uninitialized`	`create_initialized`	`create_uninitialized_and_drop`	`create_initialized_and_drop`
`large`	`66.06 ms` (🚀 2.32x faster)	`70.78 ms` (🚀 2.20x faster)	`68.71 ms` (🚀 2.24x faster)	`80.86 ms` (🚀 2.08x faster)
`medium`	`18.34 ms` (🚀 2.21x faster)	`20.73 ms` (🚀 2.07x faster)	`18.33 ms` (🚀 2.20x faster)	`36.35 ms` (✅ 1.51x faster)
`small`	`2.30 ms` (✅ 1.55x faster)	`4.28 ms` (✅ 1.29x faster)	`2.39 ms` (✅ 1.53x faster)	`21.46 ms` (❌ 1.24x slower)
`default`	`408.44 µs` (✅ 1.20x faster)	`2.19 ms` (✅ 1.05x slower)	`416.68 µs` (✅ 1.29x faster)	`21.34 ms` (❌ 1.88x slower)

shared_memory

	`copy_from_slice`	`fill`	`copy_to_slice`
`64MB`	`6.08 ms` (🚀 1.85x faster)	`3.03 ms` (✅ 1.02x faster)	`4.73 ms` (✅ 1.17x faster)
`1MB`	`49.10 µs` (✅ 1.04x slower)	`29.03 µs` (✅ 1.25x faster)	`44.40 µs` (✅ 1.11x slower)

snapshots

	`restore`	`create`
`medium`	`15.95 µs` (✅ 1.04x faster)	`44.43 ms` (✅ 1.68x faster)
`large`	`67.05 µs` (🚀 7.91x faster)	`173.59 ms` (✅ 1.70x faster)
`small`	`11.79 µs` (✅ 1.03x slower)	`2.71 ms` (✅ 1.46x faster)
`default`	`11.17 µs` (✅ 1.00x slower)	`301.70 µs` (✅ 1.27x faster)

Summary

Biggest gain: snapshots/restore/large — 🚀 7.91x faster
Worst regression: sandboxes/create_initialized_and_drop/default — ❌ 1.88x slower

mshv3 / amd (Linux) (❌ *2.19x slower* → 🚀 **2.85x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`12.21 ms` (❌ 1.77x slower)	`8.42 ms` (❌ 1.59x slower)

guest_calls

	`call_with_host_function`	`call_with_restore`	`call`	`different_thread`	`interrupt_latency`
`default`	`67.83 µs` (✅ 1.35x faster)	`154.02 µs` (✅ 1.07x faster)	`39.15 µs` (✅ 1.35x faster)
`small`	`69.99 µs` (✅ 1.30x faster)	`154.02 µs` (✅ 1.02x slower)	`41.88 µs` (✅ 1.33x faster)
`large`	`61.27 µs` (✅ 1.49x faster)	`213.11 µs` (✅ 1.09x faster)	`39.94 µs` (✅ 1.39x faster)
				`54.00 µs` (✅ 1.42x faster)	`45.40 µs` (✅ 1.09x faster)
`medium`	`69.47 µs` (✅ 1.23x faster)	`173.44 µs` (✅ 1.04x slower)	`43.01 µs` (✅ 1.32x faster)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`86.23 ms` (❌ 1.15x slower)

sample_workloads

	`24K_in_8K_out_rust`	`24K_in_8K_out_c`
	`63.91 µs` (✅ 1.33x faster)	`60.65 µs` (✅ 1.33x faster)

sandboxes

	`create_initialized_and_drop`	`create_uninitialized_and_drop`	`create_uninitialized`	`create_initialized`
`medium`	`94.72 ms` (❌ 1.52x slower)	`10.57 ms` (🚀 2.46x faster)	`10.34 ms` (🚀 2.49x faster)	`12.61 ms` (🚀 2.25x faster)
`small`	`91.72 ms` (❌ 2.19x slower)	`2.38 ms` (✅ 1.71x faster)	`2.19 ms` (✅ 1.77x faster)	`3.90 ms` (✅ 1.49x faster)
`large`	`184.48 ms` (❌ 1.28x slower)	`33.93 ms` (🚀 2.85x faster)	`33.59 ms` (🚀 2.81x faster)	`39.90 ms` (🚀 2.58x faster)
`default`	`48.76 ms` (❌ 1.41x slower)	`516.20 µs` (✅ 1.26x faster)	`463.74 µs` (✅ 1.32x faster)	`1.40 ms` (✅ 1.09x faster)

shared_memory

	`copy_from_slice`	`copy_to_slice`	`fill`
`64MB`	`5.30 ms` (✅ 1.06x faster)	`5.18 ms` (✅ 1.06x slower)	`4.11 ms` (✅ 1.10x slower)
`1MB`	`42.62 µs` (✅ 1.03x slower)	`42.48 µs` (✅ 1.01x faster)	`41.49 µs` (✅ 1.00x slower)

snapshots

	`restore`	`create`
`default`	`82.36 µs` (✅ 1.02x faster)	`358.92 µs` (✅ 1.42x faster)
`large`	`1.80 ms` (🚀 1.86x faster)	`196.62 ms` (✅ 1.22x faster)
`medium`	`94.28 µs` (✅ 1.30x faster)	`47.91 ms` (✅ 1.20x faster)
`small`	`85.34 µs` (✅ 1.03x faster)	`2.78 ms` (✅ 1.59x faster)

Summary

Biggest gain: sandboxes/create_uninitialized_and_drop/large — 🚀 2.85x faster
Worst regression: sandboxes/create_initialized_and_drop/small — ❌ 2.19x slower

mshv3 / intel (Linux) (❌ *1.56x slower* → 🚀 **1.89x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`14.56 ms` (✅ 1.05x slower)	`11.73 ms` (❌ 1.12x slower)

guest_calls

	`different_thread`	`call_with_restore`	`interrupt_latency`	`call`	`call_with_host_function`
	`87.77 µs` (✅ 1.43x faster)		`70.06 µs` (✅ 1.23x faster)
`medium`		`264.19 µs` (✅ 1.07x faster)		`78.33 µs` (✅ 1.31x faster)	`120.92 µs` (✅ 1.39x faster)
`default`		`245.26 µs` (✅ 1.13x faster)		`79.13 µs` (✅ 1.33x faster)	`121.14 µs` (✅ 1.35x faster)
`small`		`251.80 µs` (✅ 1.09x faster)		`79.76 µs` (✅ 1.34x faster)	`124.74 µs` (✅ 1.38x faster)
`large`		`348.65 µs` (✅ 1.18x faster)		`80.72 µs` (✅ 1.28x faster)	`115.04 µs` (✅ 1.52x faster)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`113.56 ms` (✅ 1.13x faster)

sample_workloads

	`24K_in_8K_out_rust`	`24K_in_8K_out_c`
	`98.22 µs` (✅ 1.47x faster)	`93.62 µs` (✅ 1.49x faster)

sandboxes

	`create_initialized_and_drop`	`create_uninitialized_and_drop`	`create_uninitialized`	`create_initialized`
`large`	`230.76 ms` (❌ 1.49x slower)	`63.71 ms` (✅ 1.66x faster)	`63.13 ms` (✅ 1.69x faster)	`70.62 ms` (✅ 1.59x faster)
`small`	`59.00 ms` (❌ 1.40x slower)	`2.56 ms` (✅ 1.76x faster)	`2.31 ms` (🚀 1.89x faster)	`5.07 ms` (✅ 1.28x faster)
`medium`	`80.48 ms` (❌ 1.27x slower)	`17.28 ms` (✅ 1.63x faster)	`17.35 ms` (✅ 1.63x faster)	`20.46 ms` (✅ 1.54x faster)
`default`	`51.96 ms` (❌ 1.56x slower)	`386.46 µs` (✅ 1.32x faster)	`355.91 µs` (✅ 1.38x faster)	`1.52 ms` (✅ 1.10x faster)

shared_memory

	`copy_to_slice`	`copy_from_slice`	`fill`
`1MB`	`41.58 µs` (✅ 1.07x slower)	`70.21 µs` (❌ 1.12x slower)	`34.91 µs` (✅ 1.08x faster)
`64MB`	`8.69 ms` (✅ 1.09x faster)	`11.42 ms` (✅ 1.01x slower)	`7.51 ms` (✅ 1.03x slower)

snapshots

	`create`	`restore`
`medium`	`42.42 ms` (✅ 1.29x faster)	`182.54 µs` (❌ 1.14x slower)
`default`	`358.88 µs` (✅ 1.27x faster)	`160.07 µs` (✅ 1.02x slower)
`small`	`4.11 ms` (✅ 1.45x faster)	`157.98 µs` (✅ 1.03x slower)
`large`	`165.26 ms` (✅ 1.27x faster)	`1.80 ms` (✅ 1.45x faster)

Summary

Biggest gain: sandboxes/create_uninitialized/small — 🚀 1.89x faster
Worst regression: sandboxes/create_initialized_and_drop/default — ❌ 1.56x slower

hyperv-ws2025 / amd (Windows) (❌ *4.96x slower* → ✅ **1.42x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`11.54 ms` (❌ 1.18x slower)	`12.55 ms` (❌ 1.46x slower)

guest_calls

	`call`	`call_with_host_function`	`call_with_restore`	`different_thread`	`interrupt_latency`
`default`	`54.08 µs` (❌ 1.42x slower)	`92.17 µs` (❌ 1.37x slower)	`208.67 µs` (❌ 1.64x slower)
`large`	`59.01 µs` (❌ 1.46x slower)	`120.51 µs` (❌ 1.78x slower)	`445.99 µs` (❌ 1.79x slower)
`medium`	`54.22 µs` (❌ 1.42x slower)	`107.02 µs` (❌ 1.60x slower)	`276.33 µs` (❌ 1.68x slower)
`small`	`53.11 µs` (❌ 1.39x slower)	`108.09 µs` (❌ 1.71x slower)	`206.86 µs` (❌ 1.39x slower)
				`84.65 µs` (✅ 1.10x slower)	`148.91 µs` (❌ 4.96x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`4.92 s` (❌ 1.49x slower)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`94.42 µs` (❌ 1.30x slower)	`90.34 µs` (❌ 1.35x slower)

sandboxes

	`create_initialized_and_drop`	`create_initialized`	`create_uninitialized_and_drop`	`create_uninitialized`
`default`	`7.06 ms` (❌ 1.41x slower)	`6.89 ms` (❌ 1.62x slower)	`1.68 ms` (❌ 1.37x slower)	`1.90 ms` (❌ 1.29x slower)
`large`	`288.00 ms` (✅ 1.17x faster)	`239.45 ms` (✅ 1.25x faster)	`393.73 ms` (✅ 1.05x slower)	`315.94 ms` (✅ 1.10x faster)
`medium`	`96.81 ms` (❌ 1.14x slower)	`88.33 ms` (❌ 1.17x slower)	`78.66 ms` (✅ 1.21x faster)	`61.84 ms` (✅ 1.42x faster)
`small`	`19.84 ms` (❌ 1.28x slower)	`13.62 ms` (✅ 1.01x slower)	`10.73 ms` (✅ 1.18x faster)	`8.73 ms` (✅ 1.37x faster)

shared_memory

	`copy_from_slice`	`copy_to_slice`	`fill`
`1MB`	`52.17 µs` (❌ 1.22x slower)	`53.45 µs` (❌ 1.14x slower)	`43.71 µs` (✅ 1.05x slower)
`64MB`	`10.77 ms` (❌ 1.29x slower)	`13.19 ms` (❌ 1.40x slower)	`6.89 ms` (❌ 1.11x slower)

snapshots

	`create`	`restore`
`default`	`975.92 µs` (❌ 1.34x slower)	`126.94 µs` (❌ 1.68x slower)
`large`	`423.02 ms` (❌ 1.12x slower)	`65.22 ms` (❌ 1.67x slower)
`medium`	`89.67 ms` (✅ 1.00x slower)	`1.12 ms` (❌ 4.23x slower)
`small`	`11.11 ms` (✅ 1.02x faster)	`136.47 µs` (❌ 1.99x slower)

Summary

Biggest gain: sandboxes/create_uninitialized/medium — ✅ 1.42x faster
Worst regression: guest_calls/interrupt_latency — ❌ 4.96x slower

hyperv-ws2025 / intel (Windows) (❌ *6.68x slower* → ✅ **1.31x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`15.57 ms` (✅ 1.06x slower)	`12.67 ms` (✅ 1.08x slower)

guest_calls

	`call`	`call_with_host_function`	`call_with_restore`	`different_thread`	`interrupt_latency`
`default`	`87.22 µs` (❌ 1.17x slower)	`123.51 µs` (✅ 1.07x slower)	`260.67 µs` (✅ 1.08x slower)
`large`	`82.18 µs` (✅ 1.06x faster)	`135.39 µs` (❌ 1.16x slower)	`790.72 µs` (❌ 1.51x slower)
`medium`	`82.22 µs` (✅ 1.00x faster)	`122.83 µs` (✅ 1.07x faster)	`351.26 µs` (❌ 1.19x slower)
`small`	`85.73 µs` (✅ 1.08x slower)	`123.63 µs` (✅ 1.01x slower)	`262.66 µs` (✅ 1.07x slower)
				`101.76 µs` (✅ 1.11x faster)	`161.41 µs` (❌ 2.44x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`5.71 s` (❌ 1.22x slower)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`105.93 µs` (✅ 1.09x faster)	`107.90 µs` (✅ 1.08x faster)

sandboxes

	`create_initialized_and_drop`	`create_initialized`	`create_uninitialized_and_drop`	`create_uninitialized`
`default`	`8.05 ms` (✅ 1.04x slower)	`6.77 ms` (❌ 1.16x slower)	`1.34 ms` (✅ 1.05x slower)	`1.03 ms` (✅ 1.09x faster)
`large`	`398.19 ms` (✅ 1.01x slower)	`313.32 ms` (✅ 1.04x faster)	`399.59 ms` (❌ 1.14x slower)	`292.09 ms` (✅ 1.05x faster)
`medium`	`104.07 ms` (✅ 1.00x slower)	`81.05 ms` (✅ 1.08x faster)	`100.29 ms` (❌ 1.15x slower)	`75.12 ms` (✅ 1.05x faster)
`small`	`21.06 ms` (✅ 1.01x slower)	`16.71 ms` (✅ 1.04x faster)	`14.14 ms` (❌ 1.16x slower)	`12.65 ms` (❌ 1.19x slower)

shared_memory

	`copy_from_slice`	`copy_to_slice`	`fill`
`1MB`	`69.43 µs` (✅ 1.04x slower)	`71.10 µs` (✅ 1.04x slower)	`35.20 µs` (✅ 1.05x slower)
`64MB`	`14.48 ms` (✅ 1.01x slower)	`19.65 ms` (❌ 1.21x slower)	`10.98 ms` (✅ 1.10x slower)

snapshots

	`create`	`restore`
`default`	`759.73 µs` (✅ 1.31x faster)	`151.57 µs` (✅ 1.00x slower)
`large`	`428.73 ms` (✅ 1.11x faster)	`76.93 ms` (❌ 1.34x slower)
`medium`	`110.64 ms` (✅ 1.10x faster)	`6.02 ms` (❌ 6.68x slower)
`small`	`14.38 ms` (✅ 1.09x faster)	`169.71 µs` (❌ 1.19x slower)

Summary

Biggest gain: snapshots/create/default — ✅ 1.31x faster
Worst regression: snapshots/restore/medium — ❌ 6.68x slower

Introduce a new internal tooling crate (hyperlight-ci) that provides: - bench subcommand: Runs criterion benchmarks in parallel via criterion-swarm. Features include: - Configurable parallelism (-j N, defaults to all P-cores) - Configurable output modes (spinner, stream, summary) - Support for pre-built binaries (--binary) to skip rebuilds - Trailing args forwarded to criterion (filter, --exact, etc.) - bench-report subcommand: Generates markdown comparison tables from criterion's target/criterion/ JSON output via criterion-markdown. Features include: - Benchmark discovery via criterion-swarm - Optional allowlist filtering via --binary or trailing args - Output to stdout This replaces ad-hoc benchmark scripting with a unified tool suitable for both local development and CI report generation. Signed-off-by: Jorge Prendes <jorge.prendes@gmail.com>

- Add cargo alias (`cargo ci`) for convenient hyperlight-ci invocation - Update dep_benchmarks workflow to use `cargo ci bench` and generate a markdown report via `cargo ci bench-report`, posting results as a PR comment per hypervisor/cpu matrix entry - Add benchmarks job to ValidatePullRequest workflow with hypervisor and cpu matrix, gated behind docs-only and build-guests checks - Grant pull-requests: write permission for PR comment posting - Simplify Justfile bench recipes to delegate to `cargo ci bench` - Update benchmarking docs to reflect the new workflow Signed-off-by: Jorge Prendes <jorge.prendes@gmail.com>

hyperlight-gh-bot · 2026-06-18T13:35:55Z

Benchmark Results

kvm / amd (Linux) (❌ *1.92x slower* → 🚀 **2.44x faster**)

function_call_serialization

	`serialize_function_call`	`deserialize_function_call`
	`8.39 ms` (❌ 1.92x slower)	`8.17 ms` (✅ 1.05x faster)

guest_calls

	`call_with_host_function`	`call_with_restore`	`call`	`different_thread`	`interrupt_latency`
`medium`	`23.61 µs` (✅ 1.06x slower)	`70.44 µs` (✅ 1.14x faster)	`12.76 µs` (✅ 1.02x slower)
`small`	`20.75 µs` (✅ 1.10x faster)	`65.94 µs` (✅ 1.08x faster)	`11.44 µs` (✅ 1.10x faster)
`default`	`24.01 µs` (✅ 1.04x slower)	`61.98 µs` (✅ 1.01x faster)	`12.75 µs` (✅ 1.01x slower)
				`9.69 ms` (❌ 1.46x slower)	`25.97 µs` (❌ 1.28x slower)
`large`	`24.42 µs` (✅ 1.06x slower)	`89.09 µs` (✅ 1.17x faster)	`12.94 µs` (✅ 1.03x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`598.13 ms` (✅ 1.07x slower)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`20.93 µs` (🚀 1.81x faster)	`22.34 µs` (✅ 1.72x faster)

sandboxes

	`create_uninitialized`	`create_initialized`	`create_uninitialized_and_drop`	`create_initialized_and_drop`
`large`	`28.89 ms` (🚀 2.25x faster)	`27.47 ms` (🚀 2.44x faster)	`29.09 ms` (🚀 2.27x faster)	`38.79 ms` (🚀 2.02x faster)
`medium`	`8.75 ms` (🚀 2.02x faster)	`9.90 ms` (🚀 1.96x faster)	`9.02 ms` (🚀 1.99x faster)	`24.61 ms` (✅ 1.25x faster)
`small`	`2.28 ms` (✅ 1.65x faster)	`3.66 ms` (✅ 1.36x faster)	`2.32 ms` (✅ 1.54x faster)	`19.58 ms` (❌ 1.25x slower)
`default`	`403.02 µs` (✅ 1.28x faster)	`1.86 ms` (✅ 1.03x slower)	`445.02 µs` (✅ 1.27x faster)	`18.91 ms` (❌ 1.63x slower)

shared_memory

	`copy_from_slice`	`fill`	`copy_to_slice`
`64MB`	`4.70 ms` (✅ 1.10x slower)	`2.96 ms` (❌ 1.18x slower)	`5.50 ms` (❌ 1.35x slower)
`1MB`	`40.14 µs` (✅ 1.02x slower)	`28.00 µs` (✅ 1.30x faster)	`41.95 µs` (❌ 1.11x slower)

snapshots

	`restore`	`create`
`medium`	`15.99 µs` (✅ 1.06x faster)	`30.21 ms` (✅ 1.16x faster)
`large`	`36.55 µs` (✅ 1.46x faster)	`98.60 ms` (✅ 1.42x faster)
`small`	`11.40 µs` (✅ 1.08x slower)	`2.88 ms` (✅ 1.22x faster)
`default`	`9.43 µs` (✅ 1.06x faster)	`305.87 µs` (✅ 1.29x faster)

Summary

Biggest gain: sandboxes/create_initialized/large — 🚀 2.44x faster
Worst regression: function_call_serialization/serialize_function_call — ❌ 1.92x slower

kvm / intel (Linux) (❌ *1.70x slower* → 🚀 **5.39x faster**)

function_call_serialization

	`serialize_function_call`	`deserialize_function_call`
	`7.15 ms` (❌ 1.25x slower)	`9.96 ms` (✅ 1.25x faster)

guest_calls

	`call_with_host_function`	`call_with_restore`	`call`	`different_thread`	`interrupt_latency`
`medium`	`38.21 µs` (✅ 1.01x slower)	`39.63 µs` (✅ 1.04x faster)	`19.73 µs` (✅ 1.01x slower)
`small`	`34.55 µs` (✅ 1.07x faster)	`35.37 µs` (✅ 1.06x faster)	`19.86 µs` (✅ 1.07x slower)
`default`	`38.21 µs` (✅ 1.03x slower)	`35.90 µs` (✅ 1.06x faster)	`19.70 µs` (✅ 1.03x slower)
				`10.29 ms` (❌ 1.55x slower)	`15.21 µs` (✅ 1.78x faster)
`large`	`38.51 µs` (✅ 1.09x slower)	`75.44 µs` (✅ 1.02x faster)	`18.43 µs` (✅ 1.05x faster)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`638.15 ms` (✅ 1.07x faster)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`30.93 µs` (🚀 3.40x faster)	`31.88 µs` (🚀 3.47x faster)

sandboxes

	`create_uninitialized`	`create_initialized`	`create_uninitialized_and_drop`	`create_initialized_and_drop`
`large`	`75.42 ms` (🚀 2.03x faster)	`78.01 ms` (🚀 2.00x faster)	`74.83 ms` (🚀 2.06x faster)	`91.67 ms` (🚀 1.84x faster)
`medium`	`19.55 ms` (🚀 2.08x faster)	`22.36 ms` (🚀 1.92x faster)	`19.71 ms` (🚀 2.04x faster)	`35.27 ms` (✅ 1.56x faster)
`small`	`2.21 ms` (✅ 1.61x faster)	`4.23 ms` (✅ 1.31x faster)	`2.41 ms` (✅ 1.52x faster)	`20.69 ms` (❌ 1.20x slower)
`default`	`376.93 µs` (✅ 1.30x faster)	`2.05 ms` (✅ 1.02x faster)	`407.56 µs` (✅ 1.32x faster)	`19.27 ms` (❌ 1.70x slower)

shared_memory

	`copy_from_slice`	`fill`	`copy_to_slice`
`64MB`	`6.08 ms` (🚀 1.85x faster)	`3.39 ms` (✅ 1.10x slower)	`4.56 ms` (✅ 1.22x faster)
`1MB`	`47.28 µs` (✅ 1.00x faster)	`27.83 µs` (✅ 1.31x faster)	`44.24 µs` (✅ 1.08x slower)

snapshots

	`restore`	`create`
`medium`	`15.12 µs` (✅ 1.11x faster)	`47.89 ms` (✅ 1.56x faster)
`large`	`66.00 µs` (🚀 5.39x faster)	`193.81 ms` (✅ 1.53x faster)
`small`	`11.61 µs` (✅ 1.03x slower)	`2.72 ms` (✅ 1.46x faster)
`default`	`10.68 µs` (✅ 1.02x faster)	`299.04 µs` (✅ 1.30x faster)

Summary

Biggest gain: snapshots/restore/large — 🚀 5.39x faster
Worst regression: sandboxes/create_initialized_and_drop/default — ❌ 1.70x slower

mshv3 / amd (Linux) (❌ *1.98x slower* → 🚀 **2.60x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`13.03 ms` (❌ 1.89x slower)	`9.02 ms` (❌ 1.71x slower)

guest_calls

	`call_with_host_function`	`call_with_restore`	`call`	`different_thread`	`interrupt_latency`
`default`	`68.32 µs` (✅ 1.35x faster)	`133.63 µs` (✅ 1.22x faster)	`39.51 µs` (✅ 1.34x faster)
`small`	`66.41 µs` (✅ 1.37x faster)	`150.98 µs` (✅ 1.00x faster)	`44.17 µs` (✅ 1.25x faster)
`large`	`70.43 µs` (✅ 1.29x faster)	`208.55 µs` (✅ 1.12x faster)	`40.19 µs` (✅ 1.39x faster)
				`57.63 µs` (✅ 1.34x faster)	`44.31 µs` (✅ 1.12x faster)
`medium`	`70.11 µs` (✅ 1.28x faster)	`160.10 µs` (✅ 1.04x faster)	`42.22 µs` (✅ 1.31x faster)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`86.99 ms` (❌ 1.16x slower)

sample_workloads

	`24K_in_8K_out_rust`	`24K_in_8K_out_c`
	`57.44 µs` (✅ 1.49x faster)	`61.58 µs` (✅ 1.31x faster)

sandboxes

	`create_initialized_and_drop`	`create_uninitialized_and_drop`	`create_uninitialized`	`create_initialized`
`medium`	`106.92 ms` (❌ 1.71x slower)	`11.47 ms` (🚀 2.27x faster)	`11.05 ms` (🚀 2.32x faster)	`13.89 ms` (🚀 2.04x faster)
`small`	`82.80 ms` (❌ 1.98x slower)	`2.42 ms` (✅ 1.68x faster)	`2.17 ms` (✅ 1.79x faster)	`4.42 ms` (✅ 1.31x faster)
`large`	`198.04 ms` (❌ 1.37x slower)	`37.85 ms` (🚀 2.56x faster)	`36.37 ms` (🚀 2.60x faster)	`43.58 ms` (🚀 2.36x faster)
`default`	`49.51 ms` (❌ 1.44x slower)	`497.14 µs` (✅ 1.29x faster)	`459.41 µs` (✅ 1.33x faster)	`1.59 ms` (✅ 1.04x slower)

shared_memory

	`copy_from_slice`	`copy_to_slice`	`fill`
`64MB`	`5.35 ms` (✅ 1.05x faster)	`5.42 ms` (❌ 1.11x slower)	`4.38 ms` (❌ 1.17x slower)
`1MB`	`42.31 µs` (✅ 1.01x slower)	`43.13 µs` (✅ 1.00x faster)	`41.46 µs` (✅ 1.01x slower)

snapshots

	`restore`	`create`
`default`	`82.39 µs` (✅ 1.02x faster)	`357.16 µs` (✅ 1.43x faster)
`large`	`1.62 ms` (🚀 2.07x faster)	`195.13 ms` (✅ 1.23x faster)
`medium`	`98.24 µs` (✅ 1.18x faster)	`48.58 ms` (✅ 1.18x faster)
`small`	`87.38 µs` (✅ 1.00x slower)	`2.90 ms` (✅ 1.52x faster)

Summary

Biggest gain: sandboxes/create_uninitialized/large — 🚀 2.60x faster
Worst regression: sandboxes/create_initialized_and_drop/small — ❌ 1.98x slower

mshv3 / intel (Linux) (❌ *2.33x slower* → 🚀 **1.94x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`14.40 ms` (✅ 1.04x slower)	`11.87 ms` (❌ 1.13x slower)

guest_calls

	`call`	`call_with_restore`	`call_with_host_function`	`interrupt_latency`	`different_thread`
`small`	`79.07 µs` (✅ 1.34x faster)	`243.84 µs` (✅ 1.14x faster)	`115.85 µs` (✅ 1.43x faster)
`large`	`70.91 µs` (✅ 1.45x faster)	`331.88 µs` (✅ 1.24x faster)	`106.94 µs` (✅ 1.65x faster)
`medium`	`77.63 µs` (✅ 1.32x faster)	`236.32 µs` (✅ 1.21x faster)	`113.70 µs` (✅ 1.49x faster)
`default`	`76.25 µs` (✅ 1.39x faster)	`240.87 µs` (✅ 1.15x faster)	`118.87 µs` (✅ 1.40x faster)
				`61.04 µs` (✅ 1.41x faster)	`82.67 µs` (✅ 1.54x faster)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`114.29 ms` (✅ 1.12x faster)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`92.98 µs` (✅ 1.54x faster)	`94.16 µs` (✅ 1.53x faster)

sandboxes

	`create_uninitialized`	`create_uninitialized_and_drop`	`create_initialized_and_drop`	`create_initialized`
`medium`	`16.69 ms` (✅ 1.69x faster)	`17.05 ms` (✅ 1.65x faster)	`147.48 ms` (❌ 2.33x slower)	`20.09 ms` (✅ 1.56x faster)
`default`	`362.39 µs` (✅ 1.36x faster)	`371.26 µs` (✅ 1.39x faster)	`49.84 ms` (❌ 1.50x slower)	`1.46 ms` (✅ 1.14x faster)
`large`	`61.97 ms` (✅ 1.73x faster)	`63.88 ms` (✅ 1.66x faster)	`224.00 ms` (❌ 1.44x slower)	`69.87 ms` (✅ 1.61x faster)
`small`	`2.24 ms` (🚀 1.94x faster)	`2.90 ms` (✅ 1.55x faster)	`97.52 ms` (❌ 2.31x slower)	`5.08 ms` (✅ 1.27x faster)

shared_memory

	`fill`	`copy_from_slice`	`copy_to_slice`
`64MB`	`7.40 ms` (✅ 1.02x slower)	`11.69 ms` (✅ 1.03x slower)	`8.95 ms` (✅ 1.06x faster)
`1MB`	`33.44 µs` (✅ 1.14x faster)	`63.75 µs` (✅ 1.02x slower)	`41.50 µs` (✅ 1.07x slower)

snapshots

	`create`	`restore`
`large`	`162.87 ms` (✅ 1.29x faster)	`1.74 ms` (✅ 1.50x faster)
`default`	`354.95 µs` (✅ 1.28x faster)	`153.93 µs` (✅ 1.03x faster)
`small`	`4.39 ms` (✅ 1.36x faster)	`158.39 µs` (✅ 1.03x slower)
`medium`	`41.74 ms` (✅ 1.32x faster)	`169.35 µs` (✅ 1.04x faster)

Summary

Biggest gain: sandboxes/create_uninitialized/small — 🚀 1.94x faster
Worst regression: sandboxes/create_initialized_and_drop/medium — ❌ 2.33x slower

hyperv-ws2025 / amd (Windows) (❌ *8.66x slower* → ✅ **1.11x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`13.63 ms` (❌ 1.39x slower)	`12.67 ms` (❌ 1.48x slower)

guest_calls

	`call`	`call_with_host_function`	`call_with_restore`	`different_thread`	`interrupt_latency`
`default`	`48.65 µs` (❌ 1.27x slower)	`78.79 µs` (❌ 1.23x slower)	`174.66 µs` (❌ 1.26x slower)
`large`	`53.10 µs` (❌ 1.29x slower)	`93.00 µs` (❌ 1.43x slower)	`461.88 µs` (❌ 1.78x slower)
`medium`	`51.24 µs` (❌ 1.40x slower)	`96.78 µs` (❌ 1.44x slower)	`229.54 µs` (❌ 1.40x slower)
`small`	`50.33 µs` (❌ 1.33x slower)	`78.57 µs` (❌ 1.26x slower)	`187.71 µs` (❌ 1.30x slower)
				`78.47 µs` (✅ 1.04x slower)	`153.79 µs` (❌ 5.12x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`4.74 s` (❌ 1.44x slower)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`98.65 µs` (❌ 1.34x slower)	`82.02 µs` (❌ 1.17x slower)

sandboxes

	`create_initialized_and_drop`	`create_initialized`	`create_uninitialized_and_drop`	`create_uninitialized`
`default`	`8.24 ms` (❌ 1.64x slower)	`6.29 ms` (❌ 1.48x slower)	`1.39 ms` (❌ 1.24x slower)	`1.50 ms` (✅ 1.00x faster)
`large`	`405.55 ms` (❌ 1.21x slower)	`320.54 ms` (✅ 1.07x slower)	`388.69 ms` (✅ 1.04x slower)	`311.94 ms` (✅ 1.11x faster)
`medium`	`104.65 ms` (❌ 1.23x slower)	`107.36 ms` (❌ 1.42x slower)	`95.10 ms` (✅ 1.00x slower)	`79.46 ms` (✅ 1.10x faster)
`small`	`20.52 ms` (❌ 1.32x slower)	`17.29 ms` (❌ 1.28x slower)	`13.25 ms` (✅ 1.04x slower)	`15.18 ms` (❌ 1.27x slower)

shared_memory

	`copy_from_slice`	`copy_to_slice`	`fill`
`1MB`	`50.56 µs` (❌ 1.20x slower)	`51.72 µs` (❌ 1.15x slower)	`44.61 µs` (✅ 1.09x slower)
`64MB`	`18.59 ms` (❌ 2.23x slower)	`18.27 ms` (❌ 1.93x slower)	`12.48 ms` (❌ 2.02x slower)

snapshots

	`create`	`restore`
`default`	`872.32 µs` (❌ 1.17x slower)	`101.89 µs` (❌ 1.40x slower)
`large`	`446.46 ms` (❌ 1.18x slower)	`59.04 ms` (❌ 1.51x slower)
`medium`	`120.42 ms` (❌ 1.35x slower)	`2.29 ms` (❌ 8.66x slower)
`small`	`15.43 ms` (❌ 1.36x slower)	`115.82 µs` (❌ 1.74x slower)

Summary

Biggest gain: sandboxes/create_uninitialized/large — ✅ 1.11x faster
Worst regression: snapshots/restore/medium — ❌ 8.66x slower

hyperv-ws2025 / intel (Windows) (❌ *3.32x slower* → ✅ **1.24x faster**)

function_call_serialization

	`deserialize_function_call`	`serialize_function_call`
	`16.46 ms` (❌ 1.12x slower)	`13.95 ms` (❌ 1.19x slower)

guest_calls

	`call`	`call_with_host_function`	`call_with_restore`	`different_thread`	`interrupt_latency`
`default`	`77.41 µs` (✅ 1.04x slower)	`130.89 µs` (✅ 1.08x slower)	`234.50 µs` (✅ 1.05x faster)
`large`	`89.57 µs` (✅ 1.01x slower)	`127.20 µs` (❌ 1.19x slower)	`784.62 µs` (❌ 1.38x slower)
`medium`	`79.85 µs` (✅ 1.03x faster)	`139.73 µs` (✅ 1.05x slower)	`353.80 µs` (❌ 1.19x slower)
`small`	`79.49 µs` (✅ 1.02x faster)	`134.30 µs` (✅ 1.08x slower)	`254.65 µs` (✅ 1.06x slower)
				`100.42 µs` (✅ 1.14x faster)	`148.20 µs` (❌ 2.24x slower)

guest_functions_with_large_parameters

	`guest_call_with_large_parameters`
	`4.71 s` (✅ 1.01x slower)

sample_workloads

	`24K_in_8K_out_c`	`24K_in_8K_out_rust`
	`97.65 µs` (✅ 1.24x faster)	`97.15 µs` (✅ 1.23x faster)

sandboxes

	`create_initialized_and_drop`	`create_initialized`	`create_uninitialized_and_drop`	`create_uninitialized`
`default`	`7.45 ms` (✅ 1.04x faster)	`6.10 ms` (✅ 1.04x slower)	`1.60 ms` (❌ 1.31x slower)	`1.42 ms` (❌ 1.28x slower)
`large`	`383.32 ms` (✅ 1.02x faster)	`320.47 ms` (✅ 1.02x faster)	`408.89 ms` (❌ 1.17x slower)	`301.74 ms` (✅ 1.02x faster)
`medium`	`111.60 ms` (✅ 1.07x slower)	`84.30 ms` (✅ 1.03x faster)	`97.63 ms` (❌ 1.12x slower)	`73.95 ms` (✅ 1.07x faster)
`small`	`20.50 ms` (✅ 1.01x faster)	`17.24 ms` (✅ 1.01x faster)	`14.45 ms` (❌ 1.18x slower)	`10.59 ms` (✅ 1.01x faster)

shared_memory

	`copy_from_slice`	`copy_to_slice`	`fill`
`1MB`	`70.83 µs` (❌ 1.16x slower)	`82.24 µs` (❌ 1.16x slower)	`37.47 µs` (❌ 1.15x slower)
`64MB`	`14.75 ms` (✅ 1.03x slower)	`17.81 ms` (✅ 1.10x slower)	`11.50 ms` (❌ 1.15x slower)

snapshots

	`create`	`restore`
`default`	`810.67 µs` (✅ 1.24x faster)	`134.78 µs` (✅ 1.11x faster)
`large`	`433.71 ms` (✅ 1.10x faster)	`83.49 ms` (❌ 1.45x slower)
`medium`	`112.00 ms` (✅ 1.09x faster)	`2.99 ms` (❌ 3.32x slower)
`small`	`14.84 ms` (✅ 1.06x faster)	`161.35 µs` (❌ 1.11x slower)

Summary

Biggest gain: snapshots/create/default — ✅ 1.24x faster
Worst regression: snapshots/restore/medium — ❌ 3.32x slower

jprendes force-pushed the benchmark-prs branch 2 times, most recently from 8ad9393 to 0e181c4 Compare June 12, 2026 10:43

jprendes added the kind/enhancement For PRs adding features, improving functionality, docs, tests, etc. label Jun 12, 2026

jprendes force-pushed the benchmark-prs branch 4 times, most recently from bc499ea to a277ac1 Compare June 17, 2026 14:48

jsturtevant mentioned this pull request Jun 17, 2026

Make Benchmarks Visible #1496

Open

jprendes force-pushed the benchmark-prs branch 3 times, most recently from e660953 to 9ce9ed0 Compare June 18, 2026 11:22

jprendes added 2 commits June 18, 2026 14:15

jprendes force-pushed the benchmark-prs branch from 9ce9ed0 to 6abf01f Compare June 18, 2026 13:15

hyperlight-dev deleted a comment from hyperlight-gh-bot Bot Jun 18, 2026

Conversation

jprendes commented Jun 12, 2026

Uh oh!

hyperlight-gh-bot Bot commented Jun 18, 2026

Benchmark Results

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

Uh oh!

hyperlight-gh-bot Bot commented Jun 18, 2026

Benchmark Results

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary

function_call_serialization

guest_calls

guest_functions_with_large_parameters

sample_workloads

sandboxes

shared_memory

snapshots

Summary