We propose a new method for rapid and reliable data delivery called R2D2. R2D2 exploits the homogeneity present in data center topologies to aggregate L3 flows into one L2 meta-flow. The simplicity of R2D2 makes it a cost-effective method for improving interconnection fabric reliability. We tested a Linux prototype of R2D2 in 1GbE and 10GbE environments and different workloads, and find that it improves TCP performance significantly.
Generates Incast by sending synchronous requests to multiple servers.
Download