Load Balancing Across Links – Detailed Explanation
1. What is Load Balancing in Networking?
Load balancing is the practice of distributing network traffic across multiple links or paths to make the best use of available bandwidth, ensure redundancy, and improve reliability.
- Prevents any one link from being overloaded.
- Increases network fault tolerance – if one link fails, others carry the load.
- Optimizes performance across the network.
2. Types of Load Balancing
| Type | Description | Notes |
|---|---|---|
| Per-Packet | Each packet is sent over a different link in sequence (round-robin). | Can cause out-of-order packets (not ideal for TCP/VoIP). |
| Per-Flow | All packets in a flow (e.g., TCP session) use the same link, but different flows use different links. | Maintains order, preferred for most enterprise uses. |
| Per-Destination | Packets are distributed based on destination address. | Simple, but less granular than per-flow. |
3. Load Balancing in Layer 2 Networks
Switches use link aggregation (such as EtherChannel, LACP, PAgP) to combine multiple links into a single logical interface.
Traffic is distributed based on MAC address, IP address, or Layer 4 port information.
4. EtherChannel and Link Aggregation Overview
- EtherChannel (Cisco): Bundles 2–8 (or more) links into a Port-Channel interface.
- LACP (IEEE 802.3ad): Open standard for dynamic aggregation.
- Purpose: Increases bandwidth and redundancy without blocking links in STP.
5. Load Balancing Criteria in EtherChannel
| Criteria | How Traffic is Distributed |
|---|---|
| Source MAC | Based on the sender’s MAC address |
| Destination MAC | Based on the receiver’s MAC address |
| Source/Destination IP | Uses source and destination IP addresses (default in many modern switches) |
| Layer 4 (Port) | Uses TCP/UDP port numbers for more granular balancing |
| Combinations | Hashing multiple fields for best distribution |
Example: If EtherChannel is set to hash based on source/destination IP, all traffic between
192.168.1.10 and 10.0.0.1 uses one link, while another flow (e.g., 192.168.1.20 to 10.0.0.2) may use another link.
6. Per-Packet vs. Per-Flow Load Balancing
- Per-Packet: Each packet may take a different path – high utilization but risk of out-of-order delivery.
- Per-Flow: All packets for a session use the same link – keeps order, best for TCP/VoIP.
7. Load Balancing in Layer 3 Networks
Equal-Cost Multipath Routing (ECMP): Routers can install multiple next-hop paths to the same destination, if the metric is identical. Traffic is distributed among them.
Example: If OSPF learns two equal-cost paths to
10.0.0.0/8, the router uses both for outbound traffic, balancing flows across interfaces.
8. Load Balancing in Dynamic Routing Protocols
| Protocol | Load Balancing | Key Feature |
|---|---|---|
| OSPF | ECMP by default (up to 4/16 paths) | Equal-cost only |
| EIGRP | ECMP and Unequal-cost (via variance) | Unique “variance” allows use of lower-priority paths |
| BGP | Supports ECMP in modern versions | Multiple paths for same prefix (if enabled) |
OSPF:
maximum-paths 4
EIGRP:
maximum-paths 4
variance 2
9. Load Balancing in Link Aggregation Protocols
PAgP and LACP both use hash algorithms (MAC/IP/Port-based) to distribute traffic.
By default, most use per-flow balancing for better packet order.
10. Hardware and Software Load Balancing
- ASIC-based forwarding: Most switches/routers use hardware for fast calculations and traffic splitting.
- Software-based: Slower, used on low-end devices or for special cases.
11. Limitations and Challenges
- Hash Collisions: Flows with same hash all use the same link, causing imbalance.
- Asymmetric Routing: Return paths may be different than outbound, complicating troubleshooting.
- Flow Stickiness: Heavy flows (like backups) can dominate a link even with multiple available.
12. Monitoring and Troubleshooting Load Balancing
# EtherChannel hashing method
show etherchannel load-balance
# Port-Channel status
show etherchannel summary
show interfaces port-channel [number]
# Routing Table (ECMP)
show ip route
13. Best Practices for Load Balancing
- Pick the right hashing method (IP-based for Layer 3, MAC-based for Layer 2).
- Make sure both sides of the link aggregation use matching settings.
- Regularly monitor bandwidth usage and adjust as needed.
- For ECMP, keep path costs identical for effective balancing.
14. Advanced Topics
- Data Center Fabrics (SDN): Use ECMP and dynamic, flowlet-based load balancing for huge scalability.
- Vendor Innovations: Some platforms support adaptive algorithms for real-time flow redistribution.
EtherChannel Example:
Switch A and Switch B are connected with four 1 Gbps links as EtherChannel (Port-Channel1).
Load balancing uses source-destination IP. Traffic from PC1 to Server1 may use one link; PC2 to Server1 may use another, depending on hash.
Layer 3 ECMP Example:
Router R1 has two equal-cost OSPF paths to 10.10.10.0/24. Both outbound interfaces carry traffic for different flows.
Switch A and Switch B are connected with four 1 Gbps links as EtherChannel (Port-Channel1).
Load balancing uses source-destination IP. Traffic from PC1 to Server1 may use one link; PC2 to Server1 may use another, depending on hash.
Layer 3 ECMP Example:
Router R1 has two equal-cost OSPF paths to 10.10.10.0/24. Both outbound interfaces carry traffic for different flows.
Summary Table: Load Balancing Key Points
| Aspect | Key Detail |
|---|---|
| Technique | EtherChannel, LACP, PAgP, ECMP (routing) |
| Common Types | Per-packet, per-flow, per-destination |
| Best Practice | Per-flow (hash-based) for stability; avoid out-of-order issues |
| Config Requirement | Consistent EtherChannel settings on both sides |
| ECMP Support | OSPF, EIGRP (equal/unequal cost), BGP |
| Monitoring | show etherchannel load-balance, show ip route |
| Typical Use Cases | Switch uplinks, core/distribution, data center fabrics |
When and Where to Use Load Balancing:
- EtherChannel: Uplinks between switches, high-traffic server connections, campus core.
- ECMP: Core routers, WAN edge, data center gateways.
- Data Centers: For SDN, virtualization, and massive throughput.