3 ms·

Depends on the design. At a minimum you're ping-ponging between userland and kernel; but you might also be bouncing between a transport layer unwrapper, an aut

Depends on the design. At a minimum you're ping-ponging between userland and kernel; but you might also be bouncing between a transport layer unwrapper, an authentication front-end, the database core, and a storage back-end.

jdsully 7y ago
Spectre may have changed things, but for a long time a user->kernel switch didn't imply a full TLB flush.
namibj 7y ago
Ideally, with io_uring(2)/DPDK/SR-IOV-NVMe, you can skip enough syscalls to drop their performance impact below 1%.