Common Load Balancing Algorithms and Their Characteristics

Load balancers can provide many load‑balancing methods, also known as scheduling methods or algorithms:

Round Robin

This method distributes incoming requests sequentially across all servers in the pool. It works best when servers have similar capacity and load, but can overload weaker servers if they receive requests they cannot handle.

Weighted Round Robin

This algorithm improves simple round robin by assigning a weight to each server based on its processing capability; servers with higher weights receive more consecutive requests, allowing stronger servers to handle a larger share of traffic.

Least Connection

Requests are sent to the server with the fewest active connections at the moment, preventing a server that handles long‑lasting connections from becoming overloaded even if it receives fewer total requests.

Least Connection Slow Start Time

When a new server joins the pool, a slow‑start period limits the number of connections it can receive, allowing the server to ramp up gradually without being overwhelmed.

Weighted Least Connection

This method combines connection count with server‑specific weights, offering a fair distribution that accounts for both current load and predefined capacity differences.

Agent Based Adaptive Balancing

The load balancer periodically collects a load metric (0‑99) from each server via an HTTP GET request. It adjusts weights dynamically based on these metrics, falling back to static weighted round robin when traffic is low or metrics are unreliable.

Fixed Weighted

Only the server with the highest weight is used unless it becomes unavailable, in which case the next highest‑weight server takes over; each server’s weight reflects its priority.

Weighted Response

Weights are calculated from the response times measured during health checks; the total response time of all servers is used to recompute individual weights roughly every 15 seconds.

Source IP Hash

This technique hashes the client’s source IP address to consistently map the client to the same backend server, eliminating the need to store client‑IP state but potentially causing uneven load distribution.