⚙️

TCP Offloading: TSO, GRO, RSS & Kernel Bypass

Complex

Free preview · 2 of 5 steps

NIC offloading: hardware taking work from the CPU

At 100 Gbps, the kernel TCP stack becomes a CPU bottleneck. Network performance depends on minimizing work per packet. Modern NICs implement hardware offloads — TSO for transmit, GRO/LRO for receive, RSS for multi-core distribution — and extreme cases bypass the kernel entirely with DPDK or XDP.

Understanding these mechanisms explains why network performance tuning often involves ethtool and IRQ affinity, not just application code.

Packets in motion — click to inspect:

1 / 2

Space to play/pause · ← → to step