Repository navigation
.NET 10 has ~1.8x higher RPS than .NET 11 preview 7 on a minimal ASP.NET Core app #132951
Description
A minimal ASP.NET Core app serves ~1.8x fewer requests/sec on .NET 11 preview 7 than on .NET 10.
Source
webbench.cs (file-based app, no project file):
#:sdk Microsoft.NET.Sdk.Web
var builder = WebApplication.CreateSlimBuilder(args);
builder.Logging.SetMinimumLevel(LogLevel.Warning);
var app = builder.Build();
app.MapGet("/", () => "Hello, World!");
app.Run();Script
run.sh:
#!/bin/bash
SDKS=${SDKS:-"net10 net11"}
MODES=${MODES:-"jit aot"}
SDK_ROOT=${SDK_ROOT:-$HOME/sdks}
OUT=${OUT:-/tmp/aspnet-repro-out}
BUILD=${BUILD:-1}
URL=http://127.0.0.1:5111
set -e
cd "$(cd "$(dirname "$0")" && pwd)"
trap 'kill $PID 2>/dev/null || true' EXIT
build() {
local props="-p:PublishAot=false"
[ "$2" = aot ] && props="-p:PublishAot=true -p:IlcInstructionSet=native"
rm -rf "$HOME/.local/share/dotnet/runfile/webbench-"*
DOTNET_ROOT="$SDK_ROOT/$1" PATH="$SDK_ROOT/$1:$PATH" \
dotnet publish webbench.cs -c Release -r linux-x64 --self-contained $props -o "$OUT/$1-$2" -v q --nologo >/dev/null
}
run() {
ASPNETCORE_URLS=$URL taskset -c 0-7 "$1/webbench" >/dev/null 2>&1 &
PID=$!
for _ in $(seq 1 100); do curl -sf -o /dev/null $URL && break; sleep 0.2; done
taskset -c 16-23 bombardier -c 32 -d 5s $URL >/dev/null 2>&1
taskset -c 16-23 bombardier -c 32 -d 10s $URL 2>&1 | awk '/Reqs\/sec/{print $2}'
kill $PID; wait $PID 2>/dev/null || true
sleep 1
}
if [ "$BUILD" = 1 ]; then
rm -rf "$OUT"; mkdir -p "$OUT"
for s in $SDKS; do for m in $MODES; do build "$s" "$m"; done; done
fi
printf "%-8s %12s %12s\n" sdk "JIT rps" "NAOT rps"
for s in $SDKS; do
j=""; a=""
for m in $MODES; do r=$(run "$OUT/$s-$m"); [ "$m" = jit ] && j=$r || a=$r; done
printf "%-8s %12s %12s\n" "$s" "$j" "$a"
doneResults
Median of 3 runs, requests/sec:
| SDK | JIT RPS | NAOT RPS |
|---|---|---|
| 10.0.11 | 326,676 | 188,161 |
| 11.0.0-preview.7.26381.103 | 177,858 | 176,568 |
- JIT regressed by 45.6% in .NET 11.
- NAOT was 42.4% slower than JIT in .NET 10.
Environment
- Linux x64 (WSL2, Ubuntu 24.04), 32 logical cores
- Server pinned to cores 0-7, load generator pinned to cores 16-23
- For faster warmup, I recommend disabling Hill Climbing and Tiered Compilation
Prerequisites:
curl -sSL -o dotnet-install.sh https://dot.net/v1/dotnet-install.sh && chmod +x dotnet-install.sh
./dotnet-install.sh --channel 10.0 --install-dir $HOME/sdks/net10 --no-path
./dotnet-install.sh --version 11.0.100-preview.7.26381.103 --install-dir $HOME/sdks/net11 --no-path
curl -sSL -o ~/bin/bombardier https://github.com/codesenberg/bombardier/releases/download/v1.2.6/bombardier-linux-amd64
chmod +x ~/bin/bombardierMetadata
Metadata
Assignees
Type
Projects
- Status
Activity
dotnet-policy-service commentedon Aug 30, 2026
Tagging subscribers to this area: @agocke, @dotnet/ilc-contrib
See info in area-owners.md if you want to be subscribed.
EgorBo commentedon Aug 30, 2026
I haven't ran any investigation yet, but very likely ThreadPool/Semaphore/Spinwait related.
dotnet-policy-service commentedon Aug 30, 2026
Tagging subscribers to this area: @JulieLeeMSFT, @VSadov
See info in area-owners.md if you want to be subscribed.
[-].NET 10 has ~1.7x higher RPS than .NET 11 preview 7 on a minimal ASP.NET Core app[/-][+].NET 10 has ~1.8x higher RPS than .NET 11 preview 7 on a minimal ASP.NET Core app[/+]EgorBo commentedon Aug 30, 2026
Looks like it might be fixed by #132765 cc @VSadov
hez2010 commentedon Aug 31, 2026
Tested with the latest nightly build but it still repros.
EgorBo commentedon Aug 31, 2026
I am unable to reproduce your results.
I see no regression between .NET 10, .NET 11 P7 and latest main on Windows-x64 (kind of makes sense from the #132765 that it's mostly Unix only).
But I do see this regression fixed on linux-x64 in the latest main.
hez2010 commentedon Aug 31, 2026
It might be caused by my thread affinity config in the test script. Will test it again later.
EgorBo commentedon Aug 31, 2026
EgorBo commentedon Aug 31, 2026
@EgorBot minimalapi -linux_amd -linux_arm -aws_graviton5 -commits ae59a93,main -profiler
VSadov commentedon Sep 11, 2026
I get roughly the same results regardless net10, net11-p7, net11-rc2
This is on a relatively fast machine, in WSL.
The RPS seem lower than reported.
I also do not see AOT RPS being much lower than JIT in net10.
VSadov commentedon Sep 11, 2026
Hold on. That is because I was resolving connectivity issues to external hosts recently and left WSL network in mirrored mode.
If I turn that off I see the following:
I do see that AOT was slower in net10, everything was slower in preview7 and now everything is faster than net10.
(AOT is still behind JIT, but much less so)
EgorBo commentedon Sep 11, 2026
yep, that's more like it. I already double-confirmed locally that it's #132765
Given the backport is merged I guess we can close it.