Very slow second write to a socket

[mateuszbaran]

I am experiencing very poor performance in a client-server application that communicates over loopback when I split a single write on the server into two separate writes to a socket. I have the following code:

using Sockets

function u_server(host::IPAddr, port::Int)
    task = @async begin
        server = listen(host, port)
        while true
            sock = accept(server)
            @async begin
                while !eof(sock)
                    line = read(sock, 1)
                    t0 = time()
                    write(sock, [0x00, 0x00])
                    flush(sock)
                    t1 = time()
                    println("time: ", t1-t0)
                end
            end
        end
    end
    return task
end

function u_client(host, port::Integer)
    conn = connect(host, port)
    write(conn, 0x20)
    flush(conn)
    response_len = read(conn, 2)

    time()
    write(conn, 0x20)
    flush(conn)
    t0 = time()
    response_len = read(conn, 2)
    t1 = time()

    println("Total time (s): ", t1-t0)
    close(conn)
end

I run server by calling u_server(IPv4("127.0.0.1"), 2041) and, in a separate Julia REPL process, I run a client u_client(IPv4("127.0.0.1"), 2041). The code above has good performance, the time it takes the client to receive both bytes is on the order of 0.1ms. However, if I send the bytes separately (for example by replacing write(sock, [0x00, 0x00]) with write(sock, 0x00, 0x00)) the time measured by the client jumps to over 40ms and is consistent across multiple runs.

Version info:

Julia Version 1.1.0
Commit 80516ca202 (2019-01-21 21:24 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-4800MQ CPU @ 2.70GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, haswell)

[JeffBezanson]

This reminds me of #45 (fixed by #6302). The disable_nagle function ended up in Distributed, but it looks like it should be moved to Sockets and documented.

[mateuszbaran]

Thanks, that seems to be it. I can make a PR. By the way, do you think Nagle's algorithm should be enabled or disabled by default?

[jpsamaroo]

Nagle's algorithm is important for ensuring good performance when doing large numbers of small writes/reads to/from a TCP socket. The fact that it's on by default for TCP sockets in most (all?) Linux distributions is, to me, a signal that it's a sensible choice to leave it on by default in applications. The assumption made seems to be that high bandwidth usecases far outweight low latency usecases.