Understanding Load Balancing: Types, Architectures, and Algorithms for High-Performance Clusters

High‑performance clusters are complex because they add a task scheduler and require appropriate allocation algorithms. A load balancer acts as the task scheduler; load balancing involves more than merely equalizing compute unit load.

Classification and Architecture

Common load balancing types are DNS load, hardware load, and software load.

1. DNS Load Balancing

Definition: Resolving the same domain name to different IP addresses, typically used for geographic load balancing. For example, northern and southern users receive different addresses.

Advantages: Simple, low cost, enables nearby access and faster response.

Disadvantages: Slow updates due to DNS caching, poor scalability, limited allocation strategies.

2. Hardware Load Balancing

Definition: Dedicated hardware devices (e.g., F5, A10) that perform load balancing.

Advantages: Powerful features (multi‑layer load, many algorithms, global load), high performance (over 1 million concurrent connections), high stability, built‑in security (firewall, DDoS protection).

Disadvantages: Expensive (starting around 150 k RMB, up to millions), limited extensibility.

3. Software Load Balancing

Definition: Load balancing implemented via software such as Nginx or LVS.

Advantages: Simple deployment and maintenance, inexpensive (just a Linux server), flexible for business‑specific extensions.

Disadvantages (vs. hardware): Moderate performance (≈50 k concurrent connections for Nginx), fewer features, typically lacks firewall/DDoS protection.

4. Typical Architecture

The three mechanisms can be combined: DNS for geographic distribution, hardware for cluster‑level balancing, and software for machine‑level balancing.

Algorithms

Load‑balancing algorithms are grouped into four categories based on their goals: task‑level distribution (equal count, proportion, weight), load‑aware distribution (CPU, connections, I/O), performance‑optimal distribution (fastest response), and hash‑based distribution (consistent hashing).

1. Round Robin

Definition: Assign tasks to servers sequentially without considering server state. Simple but may ignore unhealthy servers.

2. Weighted Round Robin

Definition: Variant of round robin that uses static server weights to reflect hardware capacity. Still ignores real‑time server health.

3. Least Load First

Definition: Assign tasks to the server with the lowest current load (e.g., connections, CPU, I/O). Improves over round robin but adds complexity for load monitoring.

4. Performance‑Optimal

Definition: From the client’s perspective, direct tasks to the server with the fastest response time. Requires frequent performance measurement.

5. Hash‑Based

Definition: Compute a hash on key request attributes (source IP, session ID) and route identical hash values to the same server, useful for stateful sessions.