Skip to content

Rakaia: Scalable In-Kernel Scheduling for TCP-Based RPCs#

Netdev 0x1A (Rome, 13-16 July 2026) digest Program Adjacent pick Rome · 13–16 July 2026

Abstract Slides Recording

Rui Yang, Konstantinos Prasopoulos, Edouard Bugnion

Play the recording: Rakaia: Scalable In-Kernel Scheduling for TCP-Based RPCs

Why: Every service-mesh and RPC stack pays for the same thing: POSIX exposes byte streams, so userspace rebuilds message semantics with I/O threads, work queues and thread pools. This attacks that at the source.

What: A kernel module exposing a purely message-oriented API that hides TCP from userspace, with gRPC adapted to use it.

How: Message parsing and work-conserving scheduling happen directly in the kernel's TCP receive path, at the earliest possible point, eliminating head-of-line blocking within and across connections. Compatible with existing RPC protocols, TLS (via kTLS) and the normal TCP stack. Reported: HOL blocking eliminated across connection counts, up to 5x higher throughput-under-SLO than KCM, and 1.56x for gRPC with 23% less userspace CPU for 2% more kernel time. (From abstract only.)

Where applicable: Research prototype as an out-of-tree module; the numbers are the argument for pushing message semantics down, not something to deploy now.