Udp Socket.ReceiveFrom a lot of allocations for IPAdress

[kripergvg]

At production server, I have a simple code

mSocket.ReceiveFrom(mBuffer, SocketFlags.None, ref ep);

I call this code every millisecond. Recently I found out that almost 40% of allocations from my application are from this code. And all of the allocations are only 3 primitives: SocketAddress, IPEndPoint, IPAdress. It seems wasteful. Maybe I do something wrong? Can I read data from the UDP socket without this allocations? Can I have managed UDP transport with zero allocations?
Simple reproduction https://github.com/kripergvg/UdpAllocations
dotmemory after 4 minutes

[scalablecory]

ReceiveFrom variants have a surprisingly large number of allocations right now. This is likely to have a significant impact on QUIC. Seems worth a look at what we can do to improve things.

[scalablecory]

What do you think of adding some APIs to use SocketAddress directly:

class Socket
{
    public int SendTo(ReadOnlySpan<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);
    public ValueTask<int> SendToAsync(ReadOnlyMemory<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);

    public int ReceiveFrom(Span<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);
    public ValueTask<int> ReceiveFromAsync(Memory<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);
}

As well as some changes to conveniently access the result:

class SocketAddress
{
    public ReadOnlyMemory<byte> IPAddress { get; }
    public int Port { get; }
}

[scalablecory]

I wrote a quick proof of concept to test this.

Before:

After:

For a loop of 200,000 send/recv between 10 clients and 1 server, we remove about 2,200,000 allocations, and go from 12 GCs down to 0.

[kripergvg]

What do you think of adding some APIs to use SocketAddress directly:

class Socket
{
    public int SendTo(ReadOnlySpan<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);
    public ValueTask<int> SendToAsync(ReadOnlyMemory<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);

    public int ReceiveFrom(Span<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);
    public ValueTask<int> ReceiveFromAsync(Memory<byte> buffer, SocketFlags socketFlags, SocketAddress remoteAddress);
}

As well as some changes to conveniently access the result:

class SocketAddress
{
    public ReadOnlyMemory<byte> IPAddress { get; }
    public int Port { get; }
}

Looking amazing. It would solve my problem.

[karelz]

Duplicate of dotnet/corefx#40933