Core Concepts
Optimizing the usage of the Last Level Cache (LLC) can significantly improve the performance of the 5G User Plane Function (UPF) running on general-purpose servers.
Abstract
The paper presents 5GC2ache, a system that optimizes cache usage to improve the performance of a 5G core network's User Plane Function (UPF). The authors identify three key problems that degrade UPF performance due to incorrect LLC usage: leaky DMA, hot/cold mbuf issues, and cache contention between DDIO and CPU cores.
To address these problems, 5GC2ache has three modules:
Status Profiler: Tracks factors affecting UPF performance, including workloads, traffic, LLC stats, and DRAM bandwidth.
UPF Configurator: Conducts an offline search to find the optimal configurations for the number of RX descriptors and size of RX buffer to mitigate leaky DMA and hot/cold mbuf issues.
LLC Allocator: Dynamically adjusts the size of DDIO cache and core cache to minimize cache contention under varying traffic patterns.
The evaluation shows that 5GC2ache can increase the throughput of a commercial 5G UPF from 54.81Gbps to 76.41Gbps, a 39.41% improvement, without modifying the UPF's source code. Compared to the state-of-the-art cache optimization system ResQ, 5GC2ache achieves 29.55% higher throughput.
Stats
The throughput of the default 5G UPF configuration is 54.81Gbps.
Adopting ResQ increases the throughput to 58.98Gbps.
5GC2ache increases the throughput to 76.41Gbps, a 39.41% improvement over the default configuration and 29.55% higher than ResQ.
The RX core LLC miss rate is reduced from 63.62% in the default configuration to 5.68% in 5GC2ache.
The TX core LLC miss rate is reduced from 30.71% in the default configuration to 0.50% in 5GC2ache.
Quotes
"5GC2ache is the first system that reveals the root cause of inefficient cache usage and cache contention in 5G UPF. This analysis opens up opportunities to improve 5G UPF performance from the perspective of cache optimization."
"Compared to the default configuration of today's commercial 5G core, 5GC2ache increases the throughput of a 5G UPF from 54.81Gbps to 76.41Gbps, 39.41% improvement."
"The throughput achieved by adopting 5GC2ache is 76.41Gbps, 29.55% higher than the 58.98Gbps of ResQ."