Extending For_each_lcore for HW Topology Lcore
summary
Claim: topology-aware core selection cuts contention and improves scaling on many-core CPUs. Implementation: add a vendor-agnostic API that extends For_each_lcore (RT_get_next_lcore_extended) and optionally queries the hardware loc library to return L1/L2/L3/IO domains. Results: customers drop worker threads from ~42 to ~32, collapse 96 competing lcores into two CCX-local groups with 15 workers each, and see substantial packet-rate gains when used with pkt-distributor/event-dev/pipeline libraries.
AI-written summary · about
chapters (11)
- 0:00Introduction and team overview
- 2:28Core density and hardware trends
- 8:11Problem statement in vRAN scaling
- 12:29CDN workload challenges
- 13:11Packet core use case analysis
- 17:29AMD EPYC architecture overview
- 20:18UPF bottlenecks and solutions
- 23:56RFC proposal and API design
- 27:31LLC awareness benefits
- 30:45Smart data cache injection
- 33:15Hardware limitations and future work
related talks
20mKRCORE: A Microsecond-scale RDMA Control Plane for Elastic Computing
17mTopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed...
14mOctoSketch: Enabling Real-Time, Continuous Network Monitoring over Multiple Cores
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
14mAccelerating PyTorch FSDP Via Overlapping Collectives With In...
17mFlattened Clos: Designing High-performance Deadlock-free Expander Data Center Networks...
18mNetworking-Native RISC-V Processor for Datacenter
17mDRack: A CXL-Disaggregated Rack Architecture to Boost Inter-Rack Communication