node-replication

Node Replication library based on Black-box Concurrent Data Structures for NUMA Architectures.

This library can be used to implement a concurrent version of any single threaded data structure: It takes in a single threaded implementation of said data structure, and scales it out to multiple cores and NUMA nodes by combining three techniques: readers-writer locks, operation logging and flat combining.

How does it work

To replicate a single-threaded data structure, one needs to implement Dispatch (from node-replication). As an example, we implement Dispatch for the single-threaded HashMap from std.

use std::collections::HashMap;
use node_replication::Dispatch;

/// The node-replicated hashmap uses a std hashmap internally.
#[derive(Default)]
struct NrHashMap {
    storage: HashMap<u64, u64>,
}

/// We support mutable put operation on the hashmap.
#[derive(Clone, Debug, PartialEq)]
enum Modify {
    Put(u64, u64),
}

/// We support an immutable read operation to lookup a key from the hashmap.
#[derive(Clone, Debug, PartialEq)]
enum Access {
    Get(u64),
}

/// The Dispatch traits executes `ReadOperation` (our Access enum)
/// and `WriteOperation` (our `Modify` enum) against the replicated
/// data-structure.
impl Dispatch for NrHashMap {
    type ReadOperation = Access;
    type WriteOperation = Modify;
    type Response = Option<u64>;

    /// The `dispatch` function applies the immutable operations.
    fn dispatch(&self, op: Self::ReadOperation) -> Self::Response {
        match op {
            Access::Get(key) => self.storage.get(&key).map(|v| *v),
        }
    }

    /// The `dispatch_mut` function applies the mutable operations.
    fn dispatch_mut(&mut self, op: Self::WriteOperation) -> Self::Response {
        match op {
            Modify::Put(key, value) => self.storage.insert(key, value),
        }
    }
}

The full example (using HashMap as the underlying data-structure) can be found here. To run, execute: cargo run --example hashmap

How does it perform

The library often makes your single-threaded implementation work better than, or competitive with fine-grained locking or lock free implementations of the same data-structure.

It works especially well if

• Your data-structure exceeds the L3 cache-size of your system (you may not see any gain from replication if your data can always remain in the cache).
• All your threads need to issue mixed, mutable and immutable operations (if not alternative techniques like RCU may work better).
• You have enough DRAM to take advantage of the replication (i.e., it's typically best to use one replica per-NUMA node which means your original memory foot-print is multiplied with the amount of NUMA nodes in the system).

As an example, the following benchmark uses Rust' the hash-table with the Dispatch implementation from above (nr), and compares it against concurrent hash table implementations from crates.io (chashmap, dashmap, flurry), a HashMap protected by an RwLock (std), and urcu.

The figures show a benchmark using hash tables pre-filled with 67M entires (8 byte keys and values) and uses a uniform key distribution for operations. On the left graphs, different write ratios (0%, 10% and 80%) are shown. On the right graph, we vary the write ratio (x-axis) with 192 threads. The system has 4 NUMA nodes, so it uses 4 replicas (at x=96, a replica gets added every 24 cores). After x=96, the remaining hyper-threads are used.

Compile the library

The works with no_std and a stable rust compiler.

cargo build

If you are using a nightly rust compiler, you can compile the library to make use of some more recent features (new_uninit, and get_mut_unchecked, negative_impls):

cargo build --features unstable

As a dependency in your Cargo.toml:

node-replication = "*"

The code should currently be treated as an early release and is still work in progress. In its current form, the library is only known to work on x86 platforms (other platforms will require some changes and are untested).

Testing

There are a series of unit tests as part of the implementation and a few integration tests that check various aspects of the implementation using a stack.

You can run the tests by executing: cargo test

Benchmarks

The benchmarks (and how to execute them) are explained in more detail in the benches folder.