I'm running a server with multiple network interfaces (some are VLAN interfaces). Some of the connections are in local network with highly bursty traffic over TCP meaning they suffer at lot from having tcp_slow_start_after_idle=1. On the other hand, the interfaces that transmit data in real internet will probably induce more latency jitter if I don't have tcp_slow_start_after_idle=1.
Is there a way to configure one interface or route to have effectively setting tcp_slow_start_after_idle=0, and another interface or route to have effectively the setting tcp_slow_start_after_idle=1 at the same time?
Having a single toggle (/proc/sys/net/ipv4/tcp_slow_start_after_idle) for the whole system seems a bit heavy handed.
Related:
- https://lkml.org/lkml/2010/4/9/427
- https://lkml.iu.edu/hypermail/linux/kernel/1111.1/02240.html
- How to apply `fq_codel` to balance traffic in multiple VLAN connections?
Update: According to https://blog.donatas.net/blog/2015/08/08/slow-start-after-idle/ it seems that the Linux default tcp_slow_start_after_idle=1 and tcp_no_metrics_save=0 actually results in the worst possible throughput.
