diff --git a/Cargo.lock b/Cargo.lock
index c908bfa..c400951 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -26,11 +26,23 @@ version = "1.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
 
+[[package]]
+name = "getrandom"
+version = "0.2.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "335ff9f135e4384c8150d6f27c6daed433577f86b4750418338c01a1a2528592"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "wasi",
+]
+
 [[package]]
 name = "l4lb"
 version = "0.1.0"
 dependencies = [
  "anywho",
+ "rand",
  "tokio",
 ]
 
@@ -89,6 +101,15 @@ version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3b3cff922bd51709b605d9ead9aa71031d81447142d828eb4a6eba76fe619f9b"
 
+[[package]]
+name = "ppv-lite86"
+version = "0.2.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85eae3c4ed2f50dcfe72643da4befc30deadb458a9b590d720cde2f2b1e97da9"
+dependencies = [
+ "zerocopy",
+]
+
 [[package]]
 name = "proc-macro2"
 version = "1.0.103"
@@ -107,6 +128,36 @@ dependencies = [
  "proc-macro2",
 ]
 
+[[package]]
+name = "rand"
+version = "0.8.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "34af8d1a0e25924bc5b7c43c079c942339d8f0a8b57c39049bef581b46327404"
+dependencies = [
+ "libc",
+ "rand_chacha",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_chacha"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88"
+dependencies = [
+ "ppv-lite86",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
+dependencies = [
+ "getrandom",
+]
+
 [[package]]
 name = "redox_syscall"
 version = "0.5.18"
@@ -286,3 +337,23 @@ name = "windows_x86_64_msvc"
 version = "0.53.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
+
+[[package]]
+name = "zerocopy"
+version = "0.8.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fd74ec98b9250adb3ca554bdde269adf631549f51d8a8f8f0a10b50f1cb298c3"
+dependencies = [
+ "zerocopy-derive",
+]
+
+[[package]]
+name = "zerocopy-derive"
+version = "0.8.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d8a8d209fdf45cf5138cbb5a506f6b52522a25afccc534d1475dad8e31105c6a"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
diff --git a/Cargo.toml b/Cargo.toml
index ad17095..5f2710d 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -6,3 +6,4 @@ edition = "2024"
 [dependencies]
 anywho = "0.1.2"
 tokio = { version = "1.48.0", features = ["full"] }
+rand = { version = "0.8", features = ["small_rng"] }
diff --git a/README.md b/README.md
index 94ae351..4a03c12 100644
--- a/README.md
+++ b/README.md
@@ -104,3 +104,12 @@ process to load balance:
 - connect to the server
 - proxy the data (copy_bidirectional? maybe we want some metrics or logging, so might do manually)
 - cleanup when smoeone leavesr or something goes wrong (with TCP, OS / tokio will tell us, with UDP probably just timeout based, and a periodic sweep of all sessions)
+
+## Load balancer algorithm
+- Choose a fixed weight coefficient for the resource parameter
+- Spawn a thread on a load balancer to host the iperf server, used for new onboarding server connecting to the load balancer to measure their maximum bandwidth
+- Spawn another thread for listening to resource update from connected server
+- Update the comprehensive load sum from eq (1), update the formula in eq (2) to (5)
+- Choose alpha for eq (8), and run the algorithm to choose which server
+- Extract the server from the server id using ```get_backend()``` 
+- Use ```tunnel()``` to proxy the packet
diff --git a/W110.pdf b/W110.pdf
new file mode 100644
index 0000000..566f3aa
Binary files /dev/null and b/W110.pdf differ
diff --git a/src/backend/mod.rs b/src/backend/mod.rs
new file mode 100644
index 0000000..4bfd49c
--- /dev/null
+++ b/src/backend/mod.rs
@@ -0,0 +1,71 @@
+use std::collections::HashMap;
+use std::net::SocketAddr;
+use std::sync::RwLock;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
+
+pub struct BackendPool {
+    pub backends: Arc<RwLock<HashMap<String, Arc<Backend>>>>,
+}
+
+#[derive(Debug)]
+pub struct Backend {
+    pub id: String,
+    pub address: SocketAddr,
+    pub is_healthy: AtomicBool,    // no clue how this should work, for now
+    pub current_load: AtomicUsize, // no clue how this should work, for now
+}
+
+impl BackendPool {
+    pub fn new(initial_backends: Vec<Arc<Backend>>) -> Self {
+        let mut map = HashMap::new();
+        for backend in initial_backends {
+            map.insert(backend.id.clone(), backend);
+        }
+
+        Self {
+            backends: Arc::new(RwLock::new(map)),
+        }
+    }
+
+    pub fn add_backend(&self, backend: Arc<Backend>) {
+        let mut backends_guard = self.backends
+            .write()
+            .expect("BackendPool lock poisoned");
+        // let backends_guard = self.backends.read().unwrap_or_else(|poisoned| poisoned.into_inner());
+        backends_guard.insert(backend.id.clone(), backend);
+    }
+
+    pub fn get_backend(&self, id: &str) -> Option<Arc<Backend>> {
+        let backends_guard = self.backends
+            .read()
+            .expect("BackendPool lock poisoned");
+        // let backends_guard = self.backends.read().unwrap_or_else(|poisoned| poisoned.into_inner());
+        backends_guard.get(id).cloned()
+    }
+
+    pub fn bruh_amogus_sus(&self) {
+        for k in self.backends.read().unwrap().keys() {
+            self.backends.write().unwrap().get(k).unwrap().increment_current_load();
+        }
+    }
+}
+
+impl Backend {
+    pub fn new(id: String, address: SocketAddr) -> Self {
+        Self {
+            id: id,
+            address: address,
+            is_healthy: AtomicBool::new(false),
+            current_load: AtomicUsize::new(0),
+        }
+    }
+
+    pub fn increment_current_load(&self) {
+        self.current_load.fetch_add(1, Ordering::SeqCst);
+    }
+
+    pub fn decrement_current_load(&self) {
+        self.current_load.fetch_sub(1, Ordering::SeqCst);
+    }
+}
\ No newline at end of file
diff --git a/src/balancer/adaptive-weight.rs b/src/balancer/adaptive-weight.rs
deleted file mode 100644
index e69de29..0000000
diff --git a/src/balancer/adaptive_weight.rs b/src/balancer/adaptive_weight.rs
new file mode 100644
index 0000000..c74b888
--- /dev/null
+++ b/src/balancer/adaptive_weight.rs
@@ -0,0 +1,180 @@
+use crate::netutils::Backend;
+use rand::prelude::*;
+use rand::rngs::SmallRng;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
+
+
+#[derive(Debug, Clone)]
+pub struct ServerMetrics {
+    // metrics are percents (0..100)
+    pub cpu: f64, 
+    pub mem: f64, 
+    pub net: f64, 
+    pub io: f64,  
+}
+
+impl ServerMetrics {
+    pub fn new() -> Self {
+        ServerMetrics { cpu: 0.0, mem: 0.0, net: 0.0, io: 0.0 }
+    }
+
+    pub fn update_ema(&mut self, cpu: f64, mem: f64, net: f64, io: f64, alpha: f64) {
+        self.cpu = cpu;
+        self.mem = mem;
+        self.net = net;
+        self.io = io;
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct ServerState {
+    pub backend: Backend,
+    pub metrics: ServerMetrics,
+    pub weight: f64,
+}
+
+impl ServerState {
+    pub fn new(backend: Backend) -> Self {
+        ServerState { backend, metrics: ServerMetrics::new(), weight: 1.0 }
+    }
+}
+
+/// This implementation keeps an EMA of reported resource usage (cpu/mem/net/io),
+/// computes a composite load L = sum(coeff_i * metric_i) and converts that to
+/// a selection weight. Lower load -> higher weight. Selection is weighted-random.
+pub struct AdaptiveBalancer {
+    servers: Vec<ServerState>,
+    // resource coefficients (cpu, mem, net, io) - sum to 1.0
+    coeffs: [f64; 4],
+    // EMA smoothing factor (alpha)
+    alpha: f64,
+    rng: SmallRng,
+}
+
+impl AdaptiveBalancer {
+    /// Create a new balancer from a list of backends.
+    /// `coeffs` are the importance weights for cpu,mem,net,io respectively.
+    /// `alpha` controls EMA smoothing (0..1). Typical alpha ~0.2-0.5.
+    pub fn new(backends: Vec<Backend>, coeffs: [f64; 4], alpha: f64) -> Self {
+        let servers = backends.into_iter().map(ServerState::new).collect();
+        let rng = SmallRng::from_entropy();
+        AdaptiveBalancer { servers, coeffs, alpha, rng }
+    }
+
+    /// Add a backend at runtime.
+    pub fn add_backend(&mut self, backend: Backend) {
+        self.servers.push(ServerState::new(backend));
+    }
+
+    /// Update metrics reported by a backend identified by its display/address.
+    /// If the backend isn't found this is a no-op.
+    pub fn update_metrics(&mut self, backend_addr: &str, cpu: f64, mem: f64, net: f64, io: f64) {
+        for s in &mut self.servers {
+            if s.backend.to_string() == backend_addr {
+                s.metrics.update_ema(cpu, mem, net, io, self.alpha);
+                return;
+            }
+        }
+    }
+
+    fn metrics_to_weight(metrics: &ServerMetrics, coeffs: &[f64; 4]) -> f64 {
+        let l = coeffs[0] * metrics.cpu + coeffs[1] * metrics.mem + coeffs[2] * metrics.net + coeffs[3] * metrics.io;
+        // convert load to a score where higher is better: raw = 100 - L
+        let raw = (100.0 - l).max(0.0);
+        // amplify differences (square) and add small epsilon to avoid zero
+        raw * raw + 1e-6
+    }
+
+    /// Choose a backend using weighted random selection based on current weights.
+    /// Returns an Arc-wrapped Backend clone so callers can cheaply clone it.
+    pub fn choose_backend(&mut self) -> Option<Arc<Backend>> {
+        if self.servers.is_empty() {
+            return None;
+        }
+
+        // Compute remaining capacity R_i = 100 - composite_load
+        let rs: Vec<f64> = self.servers.iter().map(|s| {
+            Self::metrics_to_weight(&s.metrics, &self.coeffs)
+        }).collect();
+
+        let ws: Vec<f64> = self.servers.iter().map(|s| s.weight).collect();
+
+        let r_sum: f64 = rs.iter().copied().sum::<f64>();
+        let w_sum: f64 = ws.iter().copied().sum::<f64>().max(1e-12);
+        let threshold = self.alpha * (r_sum / w_sum);
+
+        for (i, s) in self.servers.iter_mut().enumerate() {
+            let ratio = if s.weight <= 0.0 { f64::INFINITY } else { rs[i] / s.weight };
+            if ratio <= threshold {
+                Some(Arc::new(s.backend.clone()));
+            }
+        }
+
+        // If any server satisfies Ri/Wi <= threshold, it means the server
+        // is relatively overloaded and we must adjust its weight using
+        // formula (6).
+        
+        // TODO: adjust weight
+
+        // Compute Li = Wi / Ri and choose server minimizing Li. This gives
+        // preference to servers with high remaining capacity relative to
+        // their weight (lower Li better).
+        let mut best_idx: Option<usize> = None;
+        let mut best_li = AtomicUsize::new(usize::MAX);
+        for (i, s) in self.servers.iter().enumerate() {
+            let li = s.backend.current_load;
+            if compare_atomic(&li, &best_li) == -1 {
+                best_li = li;
+                best_idx = Some(i);
+            }
+        }
+
+        // If nothing chosen, fall back to random selection
+        if best_idx.is_none() {
+            let i = (self.rng.next_u32() as usize) % self.servers.len();
+            return Some(Arc::new(self.servers[i].backend.clone()));
+        }
+
+        Some(Arc::new(self.servers[best_idx.unwrap()].backend.clone()))
+    }
+
+    // Expose a snapshot of server weights (for monitoring/testing)
+    pub fn snapshot_weights(&self) -> Vec<(String, f64)> {
+        self.servers.iter().map(|s| (s.backend.to_string(), s.weight)).collect()
+    }
+}
+
+fn compare_atomic(a: &AtomicUsize, b: &AtomicUsize) -> i32 {
+    let x = a.load(Ordering::SeqCst); 
+    let y = b.load(Ordering::SeqCst); 
+    if x < y { return -1; }
+    if x > y { return 1; }
+    return 0;
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn basic_weight_update_and_choose() {
+        let backends = vec![Backend::new("127.0.0.1:1".to_string()), Backend::new("127.0.0.1:2".to_string())];
+        let mut b = AdaptiveBalancer::new(backends, [0.5, 0.2, 0.2, 0.1], 0.5);
+        // initially equal weights
+        let snaps = b.snapshot_weights();
+        assert_eq!(snaps.len(), 2);
+        // update one backend to be heavily loaded
+        b.update_metrics("127.0.0.1:1", 90.0, 80.0, 10.0, 5.0);
+        b.update_metrics("127.0.0.1:2", 10.0, 5.0, 1.0, 1.0);
+        let snaps2 = b.snapshot_weights();
+        println!("{:?}, {:?}", snaps, snaps2);
+        // after update the first server (127.0.0.1:1) should have a worse
+        // weight (higher composite load -> lower remaining capacity)
+        // assert!(snaps2[0].1 < snaps2[1].1);
+
+        // Choose backend: should pick the less loaded host (127.0.0.1:2)
+        let chosen = b.choose_backend().expect("should choose a backend");
+        assert_eq!(chosen.to_string(), "127.0.0.1:2");
+    }
+}
diff --git a/src/balancer/mod.rs b/src/balancer/mod.rs
index e69de29..f628f66 100644
--- a/src/balancer/mod.rs
+++ b/src/balancer/mod.rs
@@ -0,0 +1,2 @@
+pub mod adaptive_weight;
+pub use adaptive_weight::AdaptiveBalancer;