In modern host network stacks, Intel's Data Direct I/O (DDIO) technology plays a crucial role by allowing received packet data to be accessed directly from the last-level cache (LLC) instead of host memory, accelerating packet processing. However, DDI...
In modern host network stacks, Intel's Data Direct I/O (DDIO) technology plays a crucial role by allowing received packet data to be accessed directly from the last-level cache (LLC) instead of host memory, accelerating packet processing. However, DDIO performance is often constrained by the "leaky DMA" problem, where packet data is evicted from the LLC before processing completes due to the limited capacity of the DDIO-reserved portion.
This study revisits this issue and experimentally reveals that the primary contributor to LLC misses during DDIO is not capacity limitations alone, but rather 'cache set conflicts' arising from the host network stack's memory allocation patterns.
Motivated by this finding, we propose Tyche, a DDIO-aware network stack that leverages a deeper understanding of LLC behavior to avoid these set conflicts. Tyche is based on a Stride-1 conflict-free (CF) page allocation strategy and employs three key components—a Manager, an Allocator, and a Throttler—to efficiently manage packet data residency in the LLC.
Our Linux kernel prototype evaluation demonstrates that Tyche can saturate a 200Gbps link using two fewer CPU cores than the default Linux network stack, while maintaining low LLC miss rates (under 5%). Furthermore, Tyche improves total throughput per unit of core utilization by ~1.7× across various microbenchmarks and real-world applications like NVMe-over-TCP. This validates Tyche as a practical solution for maximizing CPU efficiency in modern high-speed network environments.