Answer
Composite next hop infrastructure provides optimized data structures to handle a label-per-prefix case and to help with convergence. Prior to Junos OS Release 9.5, handling label-per-prefix consumed a lot of kernel memory, restricting overall chassis scale. Now, the standard Junos OS method is to use label-per-VPN, which is very scalable.
With chained or composite next hop, memory usage is optimized in both the kernel and the Packet Forwarding Engine. However, on the I-chip this is available only for the Packet Forwarding Engine. The scaling numbers for I-chip based platforms are:
- Junos OS Releases prior to 9.5 -- 250,000 to 300,000 labels per prefix
- Junos OS Release 9.5 and later -- 600,000 labels per prefix
Data structure optimization within the Packet Forwarding Engine leads to significant savings in DRAM, with the most gains in the next-hop space and in the Layer 2 descriptors.
For more information, click Layer 3 VPNs Feature Guide.