In linux(debian 11.2, Alma 8.5) TcpStream after write_all cache to much memory

[tickbh]

In Rust 1.65.0 and x86_64-unknown-linux-gnu build, debian 11.2

use std::{time::Duration, io::Write};
use std::net::TcpStream;

fn main() {
    // now memory is 5m
    std::thread::sleep(Duration::new(5, 0));
    {
        for i in 1..10 {
            // any redis clint
            let mut tcp = TcpStream::connect("127.0.0.1:6379").unwrap();
            let data: Vec<u8> = vec![49; 26748528];
            let _ = tcp.write_all(&data);
            std::thread::sleep(Duration::new(5, 0));
        }
    }
    // now memory is 27m
    std::thread::sleep(Duration::new(500, 0));
}

I expected to see this happen: every once run
When I run multi thread, It will increase too much memory. just like

use std::thread;
use std::{time::Duration, io::Write};
use std::net::TcpStream;

fn main() {
    // now memory is 5m
    std::thread::sleep(Duration::new(5, 0));
    for thread_id in 1 .. 10 {
        thread::spawn(move || {
            {
                for i in 1..10 {
                    // any redis clint
                    let mut tcp = TcpStream::connect("127.0.0.1:6379").unwrap();
                    let data: Vec<u8> = vec![i + 49; 26748528];
                    let _ = tcp.write_all(&data);
                    std::thread::sleep(Duration::new(5, 0));
                }
                
                println!("end thread {}", thread_id);
            }
        });
    }
    println!("end this block");
    // now memory is 237m
    std::thread::sleep(Duration::new(500, 0));
}

So my project offten out of memory by this reason. Why This cache so much memory?
In windows or macos, this memory is reduce when thread close.

[the8472]

The allocator (glibc's malloc implementation by default) may be holding onto some memory. You could check with strace. Allocations of that size likely are mmap calls and if they get deallocated it would be a munmap at the same address.

[tickbh]

When I write a big vector, It will cache so much memory use, In my network actor, in normal action the system memory will be 100m, But now the memory It 6000m(6GB). How can I fix this problem, change build args or update my rust ?

[tickbh]

In windows or in macos, it has not this problem.

[Enselic]

Triage: Thank you for reporting. Can you please provide a complete example? In particular, some dummy program so that the connection to 127.0.0.1:6379 can succeed.

I'm also curious how exactly you measure memory usage. Memory usage for a linux process is a complicated subject. Please elaborate on your methodology. Thanks!