Advances, Systems and Applications
Table 2 The comparison of flowlet-level load balancing schemes
From: PDLB: Path Diversity-aware Load Balancing with adaptive granularity in data center networks
Schemes | Key design | Advantage and Disadvantage (A &D) |
|---|---|---|
LetFlow [5] | Selects the forwarding path for each packet cluster randomly at a fixed time interval. | Suitable for asymmetric networks, but routing randomly. |
Presto [8] | Uses the central controller to collect topology information. | Suitable for asymmetric networks. It’s not necessary to modify the protocol stack and hardware, but the deployment is complex. |
Burst-Balancer [9] | Only manipulates a small amount of critical flowlets. | Suitable for symmetric and asymmetric topologies, but the overhead is not small. |
CONGA [6] | The switch selects the lightest congestion path for each flowlet according to the congestion information table and the flowlet table. | Suitable for asymmetric networks, but the feedback delay is too large and the scalability is poor. |
HULA [11] | The path of forwarding flowlet is the best next hop. | Solved the scalability issue of CONGA, and the overhead of forwarding tables is low, but has a herd effect. |
Luopan [12] | Samples the congestion information of some paths, and forwards the fixed-size flowlet to the lightest congestion path. | Low overhead, suitable for asymmetric networks. However, sensing partial congestion cannot guarantee global optimization. |