Understanding Load Balancing: Types, Tools, and Algorithms Explained

What is Load Balancing?

In the early stages of a website, a single machine often provides centralized services, but as traffic grows, performance and stability become challenges. Expanding by adding more machines forms a cluster, and a load balancer distributes client requests (e.g., www.taobao.com) across the cluster to achieve optimal resource usage, high throughput, low response time, and avoid overload.

Most Internet systems use server clusters—web, database, or cache clusters—behind a load‑balancing server that acts as the entry point, selecting the most suitable real server for each request.

Load Balancing Classification

Load balancing can be categorized mainly into four layers, with the most common being layer‑4 (transport) and layer‑7 (application) balancing.

Layer‑2 Load Balancing : Uses a virtual IP (VIP) while machines share the same IP but different MAC addresses; the balancer rewrites the destination MAC to forward traffic.

Layer‑3 Load Balancing : Machines have different IPs; the balancer forwards requests based on IP routing tables.

Layer‑4 Load Balancing : Operates at the transport layer (TCP/UDP), modifying IP and port information to forward traffic to application servers.

Layer‑7 Load Balancing : Operates at the application layer (HTTP, DNS, etc.), allowing routing decisions based on URL, browser type, language, and other HTTP attributes.

Common Load Balancing Tools

Hardware load balancers are high‑performance but expensive; software solutions dominate the Internet. The most widely used software load balancers are LVS , Nginx , and HAProxy .

1. LVS (Linux Virtual Server)

LVS provides high‑performance, high‑availability server clusters using Linux. It mainly implements layer‑4 load balancing.

The LVS architecture consists of three layers:

Load Balancer Layer : Front‑end director servers that hold routing tables and monitor real servers.

Server Array Layer : The pool of real servers (web, mail, DNS, etc.) connected via LAN/WAN.

Shared Storage Layer : Provides shared storage for real servers, often using NFS or cluster file systems.

2. Nginx

Nginx (engine x) is a web server that also functions as a reverse proxy, HTTP/HTTPS cache, and a layer‑7 load balancer.

Key features include modular design, high reliability, low memory consumption, hot deployment, strong concurrency (up to 50 k requests per second), and rich load‑balancing strategies.

3. HAProxy

HAProxy is an open‑source, C‑based solution offering high availability, TCP/HTTP load balancing, and proxy capabilities, suitable for high‑traffic web sites.

It primarily provides layer‑7 load balancing.

Load Balancing Algorithms

Algorithms determine how a balancer selects a real server. They are divided into static and dynamic categories.

Static Algorithms : Round Robin, Ratio, Priority.

Dynamic Algorithms : Least Connections, Fastest Response, Observed, Predictive, Dynamic Ratio‑APM, Dynamic Server Activation, QoS, ToS, Rule‑based, etc.

Round Robin : Cycles through servers sequentially; simple and efficient but unsuitable for write‑heavy workloads.

Random : Distributes requests randomly; similar pros/cons to Round Robin.

Hash : Uses a key to map requests to a specific server, ensuring the same key always lands on the same node—useful for write‑heavy caches.

Consistent Hashing : Limits impact of node failures to the affected keys, improving hit rates.

Range‑Based : Assigns key ranges to servers; easy horizontal scaling but can cause data skew.

Modulo‑Based : Uses key modulo number of servers; good for balanced distribution but hard to scale.

Dynamic (Least Connections, Fastest, Observed, Predictive, etc.) : Adjusts routing based on real‑time metrics such as CPU, I/O, network load, or historical performance.

Queue‑Based (Message Queue) : Converts requests to asynchronous messages, letting downstream nodes pull work when ready, eliminating real‑time load‑balancing pressure.

Each algorithm has specific advantages, disadvantages, and suitable scenarios, ranging from pure read‑only database loads to write‑intensive cache clusters.