[arm64] Perf_FileStream.FlushAsync benchmark hangs on Debian 11

[adamsitnik]

Reported by @carlossanlop offline. Carlos has hit this issue on Debian 11 arm64 with WSL2.

Repro:

git clone https://github.com/dotnet/performance.git
python3 ./performance/scripts/benchmarks_ci.py --architecture arm64 -f net7.0 --filter '*Perf_FileStream.FlushAsync*'

Source:

using (FileStream fileStream = new FileStream("repro.txt", FileMode.Create, FileAccess.Write, FileShare.Read, 4096, FileOptions.None))
{
    for (int i = 0; i < 1024; i++)
    {
        fileStream.WriteByte(default);

        await fileStream.FlushAsync();
    }
}

In theory it could be an IO issue, but we have not touched FileStream for a few months and I suspect that it's a runtime bug similar to #64980.

@janvorli @jkotas what would be the best way to determine the reason of the hang on Linux?

Tagging subscribers to this area: @dotnet/area-system-io
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Reported by @carlossanlop offline.

Repro:

git clone https://github.com/dotnet/performance.git
python3 ./performance/scripts/benchmarks_ci.py --architecture arm64 -f net7.0 --filter '*Perf_FileStream.FlushAsync*'

Source:

using (FileStream fileStream = new FileStream("repro.txt", FileMode.Create, FileAccess.Write, FileShare.Read, 4096, FileOptions.None))
{
    for (int i = 0; i < 1024; i++)
    {
        fileStream.WriteByte(default);

        await fileStream.FlushAsync();
    }
}

In theory it could be an IO issue, but we have not touched FileStream for a few months and I suspect that it's a runtime bug similar to #64980.

@janvorli @jkotas what would be the best way to determine the reason of the hang on Linux?

Author:	adamsitnik
Assignees:	-
Labels:	arch-arm64, area-System.IO, os-linux, untriaged
Milestone:	-

[adamsitnik]

[danmoseley]

Can you get a dump at the point of hang?

[janvorli]

Interestingly, my attempt to run the benchmarks on Ubuntu 18.04 arm64 on my Odroid N2 device (with AOT in my case) has hung in the Perf_FileStream.ReadAsync benchmark over the weekend. Unfortunately I haven't attempted to take the dump and just killed the run. I am running it again without AOT this time, so I'll watch for the same issue.

[jeffhandley]

I am hitting the same hang on my Surface Laptop 2 when using WSL2/Ubuntu 20.04. This is x64. I can repro it consistently while running only this benchmark. I'll connect with @adamsitnik to collect a dump.

[jeffhandley]

I uploaded the dump file to the location where we're posting our manual run results. And I observed a peculiar behavior:

• While plugged in, on the Windows host Performance power plan, the benchmark hangs frequently (but not every time)
• When on battery power, on the Performance power plan, I've been unable to repro the hang
• When plugged in or on battery power, on the Balanced power plan, I've been unable to repro the hang

[adamsitnik]

@janvorli I assume you were not using WSL?

[jeffhandley]

ReadAsync caused a hang on my WSL/Ubuntu 20.04 config as well. I uploaded dumps from that benchmark too.

[janvorli]

@janvorli I assume you were not using WSL?

Right, it was a physical device in my case.

[janvorli]

I just noticed that I have a hang in the ReadAsync in another benchmark run that I've started yesterday on the same device, this time in Alpine docker container.

[carlossanlop]

FYI - I hit this hang in:

• Ubuntu 20.04 x64 (Physical laptop)
• Ubuntu 20.04 arm64 (WSL)
• Debian 11 (WSL)