perf: default registered DataProtection works slower on linux compared to windows

DeagleGross commented 1 day ago

I ran the sample benchmark, where I get the DataProtector via IServiceProvider and ran multiple Protect() and Unprotect() on sample string:

// getting the data protector
var services = new ServiceCollection()
    .AddDataProtection()
    .Services.BuildServiceProvider();
_dataProtector = services.GetDataProtector("SamplePurpose");

// running the protect/unprotect
for (var i = 0; i < 1000; i++)
{
    _ = _dataProtector.Protect(LoremIpsumData);
}

The results are showing a big difference on windows against linux machines (results below). I was launching benchmarks using crank config and perf machines. I.e. crank --config https://raw.githubusercontent.com/DeagleGross/Baraholka/refs/heads/main/Benchmarks/Benchmarks/crank/data-protector.yml --scenario dataprotector --profile aspnet-perf-lin --application.framework net10.0

Frameworks infra:

Machine:

BenchmarkDotNet v0.14.0, Windows 10 (10.0.20348.2461) Intel Xeon E-2336 CPU 2.90GHz, 1 CPU, 12 logical and 6 physical cores .NET SDK 5.0.404 [C:\Program Files\dotnet\sdk] [Host] : .NET 10.0.0 (10.0.24.57009), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Benchmark run:

Toolchain=InProcessEmitToolchain MaxIterationCount=12 MinIterationCount=9 WarmupCount=1

Linux:

crank --config https://raw.githubusercontent.com/DeagleGross/Baraholka/refs/heads/main/Benchmarks/Benchmarks/crank/data-protector.yml --scenario dataprotector --profile aspnet-perf-lin --application.framework net10.0

Method	Mean	Error	StdDev	Gen0	Allocated
Protect	11.19 ms	0.027 ms	0.014 ms	156.2500	4.85 MB
Unprotect	10.03 ms	0.026 ms	0.015 ms	125.0000	3.64 MB

Windows:

crank --config https://raw.githubusercontent.com/DeagleGross/Baraholka/refs/heads/main/Benchmarks/Benchmarks/crank/data-protector.yml --scenario dataprotector --profile aspnet-perf-win --application.framework net10.0

Method	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
Protect	3.284 ms	0.0035 ms	0.0021 ms	66.4063	3.9063	3.9063	-
Unprotect	3.273 ms	0.0109 ms	0.0057 ms	74.2188	3.9063	3.9063	-

sebastienros commented 1 day ago

I would suggest some changes on the benchmark.

Use a payload that is similar to the one in antiforgery (token) instead of a lorem ipsum.
Don't call unprotect in the protect benchmark. Instead, precompute the protected value once ([Setup]) and use this value in the method.
Don't add loops in the benchmark methods, just keep it as-is and let BDN handle it, providing the time/memory stats for each call.

sebastienros commented 1 day ago

Also, to prevent dead code elimination, return the encrypted/decrypted value instead of void.

DeagleGross commented 1 day ago

thanks for the suggestions, @sebastienros ! I changed the benchmark, here are new results:

BenchmarkDotNet v0.14.0, Debian GNU/Linux 12 (bookworm) (container) Intel Xeon E-2336 CPU 2.90GHz, 1 CPU, 12 logical and 6 physical cores .NET SDK 8.0.404 [/usr/share/dotnet/sdk] [Host] : .NET 10.0.0 (10.0.24.57009), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Toolchain=InProcessEmitToolchain WarmupCount=5

Linux:

Method	Mean	Error	StdDev	Gen0	Allocated
Protect	10.792 µs	0.0149 µs	0.0139 µs	0.1221	3.91 KB
Unprotect	9.882 µs	0.0139 µs	0.0116 µs	0.1068	3.34 KB

Windows:

Method	Mean	Error	StdDev	Gen0	Allocated
Protect	3.125 µs	0.0027 µs	0.0024 µs	0.0381	1.19 KB
Unprotect	3.145 µs	0.0036 µs	0.0033 µs	0.0343	1.03 KB

DeagleGross commented 1 day ago

Here are the traces for the first run (unchanged benchmark code) traces-with-allocs.zip

dotnet / aspnetcore