We plan on installing a fast private LAN shared only by our servers via a 2nd NIC.
Basically, a server will "speak" (master) and all others will "listen" (slaves x 4).
Initially we thought about implementing a common tcp-ip algorithm between master and slaves, but after 2nd thoughts, an UDP based algorithm could be better:
- the master broadcasts packets to the whole LAN (instead of slave by slave sending).
- we would implement a simple re-send mechanism in case of missing packet (but probability of lost packets should be low)
Our question is related to our strategy relevancy:
- is the TCP protocol usually taken care of by the NIC itself (re-send mechanism), or by the Linux driver? (talking about recent NICs, say, since 2008 )
If it is the driver, we'd probably better use UDP and implement a simpler re-send mechanism, without the overhead of TCP.
It will not be handled by the NIC card itself. The NIC card does not know what happens at the IP layer and above.