In conventional networking, an application sends and receives data through socket calls handled by the operating system kernel. The kernel performs protocol processing, buffering, scheduling, and other work before passing traffic between the application and the network adapter. These mechanisms provide compatibility and isolation, but system calls, context switches, interrupts, and memory copies can become significant overhead for applications exchanging many small messages.
Kernel bypass removes the normal kernel networking stack from some or all of the performance-critical data path. An application uses a supported user-space library to interact with adapter queues or specialized communication facilities. The kernel typically still participates in initialization, resource allocation, memory protection, and device management. Bypassing the networking stack therefore does not mean eliminating the operating system or giving an application unrestricted access to physical memory.
Cloud implementations depend on the instance type, virtualized hardware capabilities, drivers, and application communication framework. Examples include user-space packet processing with the Data Plane Development Kit, or DPDK, and communication libraries that access a low-latency fabric. Kernel bypass and remote direct memory access, or RDMA, are related but not synonymous: kernel bypass describes avoiding the normal kernel data path, while RDMA describes accessing remote memory through supported hardware and transport mechanisms. Neither term should be treated as a synonym for zero-copy networking.
For AWS interface-selection questions, distinguish an elastic network interface, ENA, and EFA. An elastic network interface is a logical network interface that carries configuration such as IP addresses and security group associations; its presence alone does not imply kernel bypass. Elastic Network Adapter, or ENA, provides enhanced networking for high-throughput, low-latency IP traffic. Ordinary applications using sockets through the ENA kernel driver still use the kernel networking stack. Supported ENA configurations can also support user-space packet processing through DPDK, but that requires an explicitly configured application and driver path.
Elastic Fabric Adapter, or EFA, provides an OS-bypass communication capability for supported EC2 instances. Compatible MPI applications and supported distributed machine learning communication libraries can use EFA through the appropriate software stack, commonly including libfabric. Ordinary TCP socket traffic does not automatically become OS-bypass traffic merely because an instance has EFA. EFA's OS-bypass communication is restricted to instances within the same subnet, and deployments must satisfy its security group and software requirements.
For example, a tightly coupled simulation that repeatedly exchanges small messages between workers may benefit from EFA because communication latency and variability affect total execution time. A conventional HTTPS application usually benefits first from suitable ENA-capable instances, connection reuse, appropriate instance bandwidth, and application tuning. Choosing EFA without a compatible communication path will not accelerate ordinary web requests.
The trade-offs include more complex deployment, specialized troubleshooting, and potentially reduced visibility from host packet-capture tools that observe only the normal kernel path. Polling-based implementations can reserve substantial CPU capacity even when traffic is light. Benchmark application completion time, tail latency, CPU use, and total cost rather than assuming that bypass is always faster or cheaper. Provider network controls still apply; kernel bypass does not bypass security groups, network capacity limits, or placement constraints.
Amazon EC2