Support TCP_NODELAY

[alu]

Support TCP_NODELAY for performance improvement.

I tested performance in my environment, but that was not well.
It was caused by TCP_NODELAY being disabled.

My test code is here

#[test]
fn memcache_bench() {
    extern crate memcache;

    let mut conn = memcache::Client::new("memcache://localhost:20343").unwrap();

    conn.set_nodelay(true).unwrap();
    let start = ::std::time::Instant::now();
    for _ in 0..1000 {
        let _: Option<String> = conn.get("foo").unwrap();
    }
    let elapsed = start.elapsed();
    println!(
        "Elapsed (nodelay enabled): {}.{}",
        elapsed.as_secs(),
        elapsed.subsec_nanos() / 1_000_000
    );

    conn.set_nodelay(false).unwrap();
    let start = ::std::time::Instant::now();
    for _ in 0..1000 {
        let _: Option<String> = conn.get("foo").unwrap();
    }
    let elapsed = start.elapsed();
    println!(
        "Elapsed (nodelay disabled): {}.{}",
        elapsed.as_secs(),
        elapsed.subsec_nanos() / 1_000_000
    );
}

And results is

$ cargo test memcache_bench --release --lib -- --nocapture
    Finished release [optimized] target(s) in 0.0 secs
     Running *****-3c8e21312813a3b5

running 1 test
Elapsed (nodelay enabled): 0.47
Elapsed (nodelay disabled): 44.163
test pool::tests::memcache_bench ... ok

Thanks.

[coveralls]

Coverage remained the same at 0.0% when pulling bcabf52 on alu:master into 8a1e3e1 on aisk:master.

[aisk]

This is awesome and I think we should merge this 👍

But I think we should do a small change for this PR. The memcache::Client.set_nodelay function can be called whether there was some requests to memcached on the fly, and I think this will make some unexceptable behaviors. So I think we should just let users set this options when they initialize the memcache client, and can't set back after that.

So can you modify this PR, and add a new new_with_options method to the memcache::Client, and this method will receive a memcache::Client::Options struct to indicate this?

[alu]

@aisk Thank you for review, i changed the implement.

But memcache::connection::Connection::set_nodelay() is still public, i think it is not the best.

And there are some problems.

1. The function will be called from outside of this crate.
2. There is case that is unix and tcp connection mixed.
3. The option is not for Client, it's for connection.

So, I thinks pass options by query is more better such as Client::new("memcache://localhost:112211/?tcp_nodelay=true").

How dou you think it?

[aisk]

@alu Good idea!

[alu]

@aisk I tried it, please review again.

[aisk]

We can compare them by k == "tcp_nodelay" since Cow have implemented the Eq trait.

[alu]

Thanks, fixed it.

[aisk]

Sorry for another thing, I think we can add a simple test at the end of connection.rs like this:

#[cfg(test)]
mod tests {
    #[test]
    fn tcp_nodelay() {
        super::Connection::connect("memcache://localhost:12345?tcl_nodelay=true").unwrap();
    }
}

[alu]

@aisk Added the test, thanks.

[aisk]

Merged, thank you @alu ! I will release a new version now.

[alu]

@aisk Thank!

I did not rebase for cleanup commits, it's ok?

[aisk]

@alu No problem, I've merged this PR using squash merge.

0.7.0 is released, thanks!