Load Balancer Strategies

Intro

• balance requests across server pools
• depend on service or application type being served
• status of network and server at time of request
• server health / pre-defined condition

Strategies (algorithms)

Network Layer Algorithms

1. Round Robin

• simplest and most used
• distributed on rotation > circular list of available servers are given each request as received
• assumes all servers are same

2. Weighted Round Robin

• based on application server characteristics
• each server is assigned a weight based on its capabilities
• higher weight gets more requests
• Dynamic RR where weight is calculated in realtime

3. Least Connection

• dynamic allocation
• given to server with least number of active connections
• helps when some servers handle longer lived connections

4. Weighted Least Connection

• mix of active connections and the capablities of each server based on the weight assigned to it

5. Resource Based (Adaptive) **[Custom load]**

• Monitoring agent on each server reports its current load to load balancer
• based on the reports the queries are distributed
• used when traffic is predictable used in enterprises
• not suitable for uneven and sudden traffic

6. Resource Based (SDN Adaptive)

• knowledge from Layers 2, 3, 4 and 7 and input from SDN Controller
• statuse of server and applications running on them
• health of network infrastructure and congestion levels

7. Fixed Weighted

• Server with the highest weight receives all requests
• if the highest weight fails, all traffic is directed to the next highest weight servers
• WHY?? minimise spread and only as backup?
• might use 100/0, 80/20 weightage.

8. Weighted Response Time

• response times to regular health checks of each server used to assign weight
• determines which server receives the next request

9. Source IP Hash

• source and destination IP combined to generate an unique hash key
• even if a session gets disconnected, can regenerate and connect to same server to reconnect session

10. URL hash

• each read request served by specific server based on a set of content (object)
• improve backend cache performance of each server
• write requests are distributed evenly to all holding the object

11. Least Response Time

• least number of connection and least avg response time is selected

12. Least Bandwidth / Least Packets

Application Layer Algorithms

• based on content of requests including session cookies, HTTP header / message
• allows for intelligent distribution
• response data also used to monitor server performance and active load

1. Least Pending Requests (LPR)

• adapts to abrupt changing traffic
• monitor and efficiently distribute to most available servers
  1. Accurate load distribution
  2. Request specific distribution