diff --git a/benches/benches.rs b/benches/benches.rs
index b21fba329..d9fe6dc79 100644
--- a/benches/benches.rs
+++ b/benches/benches.rs
@@ -19,6 +19,7 @@ fn main() {
     suites::bench_raw_node(&mut c);
     suites::bench_progress(&mut c);
     suites::bench_progress_set(&mut c);
+    suites::bench_cluster(&mut c);
 
     c.final_summary();
 }
diff --git a/benches/suites/cluster.rs b/benches/suites/cluster.rs
new file mode 100644
index 000000000..ea1d650a5
--- /dev/null
+++ b/benches/suites/cluster.rs
@@ -0,0 +1,271 @@
+use criterion::Criterion;
+use criterion::{BatchSize, Bencher, BenchmarkId, Throughput};
+use raft::prelude::*;
+use raft::storage::MemStorage;
+use raft::{StateRole, INVALID_ID};
+use slog::Logger;
+use std::cell::Cell;
+use std::collections::{HashMap, VecDeque};
+use std::sync::mpsc::{self, Receiver, Sender, TryRecvError};
+use std::sync::{Arc, Mutex, RwLock};
+use std::thread;
+use std::time::{Duration, Instant};
+
+pub fn bench_cluster(c: &mut Criterion) {
+    static KB: usize = 1024;
+    // TODO: Entry with empty data will be omitted when being applied so that a client will
+    // never be notified. Maybe using a unique ID for each entry 
+    let mut test_sets = vec![1, 32, 128, 512, KB, 4 * KB, 32 * KB];
+    let mut group = c.benchmark_group("Raft::cluster");
+    for size in test_sets.drain(..) {
+        group
+            .measurement_time(Duration::from_secs(10))
+            .sample_size(10)
+            .throughput(Throughput::Bytes(size as u64))
+            .bench_with_input(
+                BenchmarkId::from_parameter(size),
+                &size,
+                |b: &mut Bencher, size| {
+                    let logger = raft::default_logger();
+                    let mut cluster = Cluster::new(3);
+                    cluster.start(&logger);
+                    thread::sleep(Duration::from_secs(1));
+                    b.iter_batched(
+                        || bench_proposal_msg(cluster.leader, cluster.leader, *size),
+                        |m| cluster.propose(m),
+                        BatchSize::SmallInput,
+                    )
+                },
+            );
+    }
+}
+
+fn bench_proposal_msg(from: u64, to: u64,  data_size: usize) -> Message {
+    let mut m = Message::default();
+    m.set_msg_type(MessageType::MsgPropose);
+    m.from = from;
+    m.to = to;
+    let mut e = Entry::default();
+    e.data = vec![0; data_size];
+    e.context = vec![0; 8];
+    m.set_entries(vec![e].into());
+    m
+}
+
+struct Cluster {
+    leader: u64, // always 1 if node_count >0
+    node_count: u64,
+    network: Network,
+    notify_center: Arc<Mutex<VecDeque<Sender<()>>>>,
+}
+
+impl Cluster {
+    fn new(node_count: u64) -> Self {
+        Cluster {
+            leader: INVALID_ID,
+            node_count: node_count,
+            network: Network::default(),
+            notify_center: Arc::new(Mutex::new(VecDeque::new())),
+        }
+    }
+
+    fn start(&mut self, logger: &Logger) {
+        let node_count = if self.node_count == 0 {
+            1
+        } else {
+            self.node_count
+        };
+        self.node_count = node_count;
+        // Make sure leader id is valid
+        let leader = if self.leader == INVALID_ID || self.leader > node_count {
+            1
+        } else {
+            self.leader
+        };
+        self.leader = leader;
+        let voters = (1..=node_count).fold(vec![], |mut sum, id| {
+            sum.push(id);
+            sum
+        });
+        let mut recvs = (1..=node_count).fold(vec![], |mut sum, id| {
+            let recv = self.network.register(id);
+            sum.push(recv);
+            sum
+        });
+        for (id, receiver) in (1..=node_count).zip(recvs.drain(..)) {
+            let mut config = Config {
+                id,
+                election_tick: 10,
+                heartbeat_tick: 1,
+                ..Default::default()
+            };
+            let conf_state = ConfState::from((voters.clone(), vec![]));
+            let store = MemStorage::new_with_conf_state(conf_state);
+            if id == leader {
+                config.election_tick = 3;
+            }
+            let raw_node =
+                RawNode::new(&config, store, logger).expect("Creating raw node must be successful");
+            let notify_center = self.notify_center.clone();
+            let n = self.network.clone();
+            thread::spawn(move || {
+                let mut node = Node {
+                    raft: raw_node,
+                    network: n,
+                    my_mailbox: receiver,
+                    notify_center,
+                };
+                let timer = Cell::new(Instant::now());
+                loop {
+                    node.handle_raft(&timer);
+                }
+            });
+        }
+    }
+
+    // Propose a Message and return a signal receiver to indicate the
+    // proposal is committed.
+    fn propose(&self, proposal: Message) {
+        let (tx, rx) = mpsc::channel();
+        {
+            let mut queue = self.notify_center.lock().unwrap();
+            queue.push_back(tx);
+        }
+        self.network.send(proposal);
+        rx.recv().expect("Proposal must be committed successfully")
+    }
+}
+
+struct Node {
+    raft: RawNode<MemStorage>,
+    my_mailbox: Receiver<Message>,
+    network: Network,
+    notify_center: Arc<Mutex<VecDeque<Sender<()>>>>,
+}
+
+impl Node {
+    fn handle_raft(&mut self, timer: &Cell<Instant>) {
+        // Tick every 10ms
+        thread::sleep(Duration::from_millis(10));
+        loop {
+            // Step raft messages.
+            match self.my_mailbox.try_recv() {
+                Ok(msg) => {
+                    self.raft.step(msg).expect("Step message must successful");
+                },
+                Err(TryRecvError::Empty) => break,
+                Err(TryRecvError::Disconnected) => return,
+            }
+        }
+        if timer.get().elapsed() >= Duration::from_millis(10) {
+            // Tick the raft.
+            self.raft.tick();
+            timer.set(Instant::now());
+        }
+        self.on_ready();
+    }
+
+    fn on_ready(&mut self) {
+        if !self.raft.has_ready() {
+            return;
+        }
+        let mut ready = self.raft.ready();
+        let store = self.raft.raft.raft_log.store.clone();
+        store
+            .wl()
+            .append(ready.entries())
+            .expect("Appending entries must successful");
+        if *ready.snapshot() != Snapshot::default() {
+            let s = ready.snapshot().clone();
+            store
+                .wl()
+                .apply_snapshot(s)
+                .expect("Applying snapshot must successful");
+        }
+        for msg in ready.messages.drain(..) {
+            self.network.send(msg);
+        }
+        if let Some(committed_entries) = ready.committed_entries.take() {
+            for entry in &committed_entries {
+                if entry.data.is_empty() || EntryType::EntryConfChange == entry.get_entry_type() {
+                    continue;
+                }
+                // Notify committed proposal to client
+                if self.raft.raft.state == StateRole::Leader {
+                    let signal = self.notify_center.lock().unwrap().pop_front().unwrap();
+                    signal.send(()).unwrap();
+                }
+            }
+            if let Some(last_committed) = committed_entries.last() {
+                let mut s = store.wl();
+                s.mut_hard_state().commit = last_committed.index;
+                s.mut_hard_state().term = last_committed.term;
+            }
+        }
+        self.raft.advance(ready);
+    }
+}
+
+#[derive(Default, Debug, PartialEq, Eq, Hash)]
+struct Connection {
+    from: u64,
+    to: u64,
+}
+
+type DelayMap = HashMap<Connection, (f64, Duration)>;
+
+#[derive(Default, Clone)]
+struct Network {
+    senders: HashMap<u64, Sender<Message>>,
+    delaym: Arc<RwLock<DelayMap>>,
+}
+
+impl Network {
+    fn register(&mut self, id: u64) -> Receiver<Message> {
+        let (tx, rx) = mpsc::channel();
+        self.senders.insert(id, tx);
+        rx
+    }
+
+    fn start(&mut self, mut mailboxes: HashMap<u64, Sender<Message>>) {
+        for (id, sender) in mailboxes.drain() {
+            let (tx, rx) = mpsc::channel::<Message>();
+            let delaym = self.delaym.clone();
+            thread::spawn(move || loop {
+                match rx.recv() {
+                    Ok(msg) => {
+                        let dm = delaym.read().unwrap();
+                        maybe_delay(&*dm, msg.from, msg.to);
+                        sender.send(msg).expect("Fail to send message");
+                    }
+                    Err(_) => return,
+                }
+            });
+            self.senders.insert(id, tx);
+        }
+    }
+
+    fn send(&self, m: Message) {
+        if let Some(sender) = self.senders.get(&m.to) {
+            // ignore error here
+            sender.send(m).expect("Fail to send message");
+        }
+    }
+
+    fn delay(&self, from: u64, to: u64, perc: f64, duration: Duration) {
+        if duration.as_nanos() > 0 {
+            let mut m = self.delaym.write().unwrap();
+            m.insert(Connection { from, to }, (perc, duration));
+        }
+    }
+}
+
+fn maybe_delay(delaym: &DelayMap, from: u64, to: u64) {
+    let (perc, duration) = delaym
+        .get(&Connection { from, to })
+        .cloned()
+        .unwrap_or((0f64, Duration::from_micros(0)));
+    if perc != 0f64 && rand::random::<f64>() <= perc {
+        thread::sleep(duration);
+    }
+}
diff --git a/benches/suites/mod.rs b/benches/suites/mod.rs
index f6e36f7ef..c6b325086 100644
--- a/benches/suites/mod.rs
+++ b/benches/suites/mod.rs
@@ -6,3 +6,5 @@ mod progress;
 pub use self::progress::*;
 mod progress_set;
 pub use self::progress_set::*;
+mod cluster;
+pub use self::cluster::*;