Kernel Bypass involves avoiding the kernel for system calls, and bringing those system calls to userspace. The most popular example is networking – instead of using the linux network kernel, a fast userspace one is important.
Solarflare sells a userspace TCP stack (OpenOnload) for use with their NICs.
Currently, the Linux Kernel can process 1M requests per minute. For more, you’ll need configuration.
A Userspace TCP stack is more efficient, but is hard to debug (Linux tools won’t work, so it’s up to the creator of the library to provide debugging tools).
A full kernel bypass has the best performance, but worst debuggability (no SSH!)
A partial kernel bypass works, where you get some extra performance off of a SolarFlare NIC by bypassing one RX queue.
To Kernel Bypass, Packet_mmap dodges the two system
calls that AF_PACKET uses, and sets up a ring buffer to
send or receive packets (much like io_uring).
Unfortunately Packet_mmap isn’t that fast, as packets
still go through the kernel.
Other networking frameworks, like DPDK, Netmap, or Snabbswitch take over the entire network card, which makes debugging harder.
Solarflare’s EF_VI supports OpenOnload, which
LD_PRELOADs network syscalls of the running program. Then
it has its own RX queue, hidden from the kernel, which, once a filter
rule is set up, packets can be routed to it, bypassing the kernel, and
getting that sweet 15M packet line speed.