LanParty千人打遊戲，用開源Router打造全亞洲最大網咖

Here is the summary and conclusion of the talk “LanParty千人打遊戲，用開源Router打造全亞洲最大網咖” from COSCUP, presented by Jen:

Video Overview

• Video URL:http://www.youtube.com/watch?v=jCZHz5l0fJ8
• Presenter: Jen (Network Architect / Tunghai University Adjunct Lecturer) [01:09]
• Context: Designing and maintaining high-speed, low-latency network infrastructure for WirForce, Asia’s largest BYOC (Bring Your Own Computer) LAN party with over 1,200 gamers connected 24/7 for four consecutive days [03:52].

Key Summary Points

1. 2024 Deployment: The Pain Points of Traditional Open-Source Firewalls

• Network Setup: 14 separate 1Gbps external lines connected via an L2 switch [08:36].
• Challenges Encountered:
  • Multi-WAN Load Balancing: Testing 14 independent WAN gateways in an office lab was nearly impossible beforehand [08:46].
  • Persistent Streaming (RTMP): Many gamers ran 24/7 live webcams/streams over TCP (RTMP). If the load balancer switched WAN links, TCP connections broke; furthermore, upload saturation on single lines choked download throughput [11:02].
  • Performance Limits: OPNsense/FreeBSD kernel forwarding struggled to maintain throughput under massive NAT and connection tables [10:24].
  • Emergency Rescue: The team had to fall back to an enterprise-grade A10 Networks 3501 CGNAT hardware appliance for the remainder of the event [12:21].

2. 2025 Deployment: Redesign with VyOS, DPDK & VPP

• Hardware & WAN Upgrade: Chunghwa Telecom provided 2 x 10Gbps dedicated optical lines [12:59]. The host server was upgraded to a dual Xeon E5-2696 v4 (44 cores / 88 threads) Dell R730xd with 256GB RAM and 3 x Mellanox ConnectX-4 Lx 25GbE NICs with NUMA affinity and PCI Passthrough [14:15].
• Architecture Shift (DPDK & FD.io VPP):
  • DPDK (Data Plane Development Kit): Bypasses the Linux kernel network stack, operating in user space via polling mode and hugepages to eliminate interrupt overhead and TLB misses [16:09].
  • FD.io VPP: Built on DPDK to provide full L2–L4 forwarding, routing, and NAT at extreme packet rates [18:00].
• NAT & Connection Tracking Benchmarks:
  • While traditional Linux conntrack and BSD firewalls suffer from table locks and L3 cache limits at high concurrent flow counts (100k flows), VPP scaled linearly and saturated the full 25Gbps line rate using only 8 CPU cores [23:36].

3. Local Traffic Offloading via Lancache

• The Problem: Simultaneous game downloads (e.g., 100GB+ per game) could easily saturate multi-gigabit WAN links [25:26].
• Implementation: Deployed Lancache using an 8 x 2TB NVMe SSD RAID-Z array (14TB usable) on a 25GbE link, redirecting client game CDN queries via internal DNS hijacking [25:42], [26:22].
• Results: Out of ~10TB total game data requested during the event, ~7TB was served directly from the local Lancache array, offloading over 70% of external bandwidth [26:58].

Conclusion & Future Directions

1. Kernel Bypass is Critical for Massive Multi-Gigabit NAT: Traditional OS kernel network stacks bottleneck under high-concurrency LAN party loads. Combining VyOS + DPDK/FD.io VPP proves that commodity x86 servers can replace multi-million-dollar proprietary CGNAT appliances and deliver 10G–25G line-rate performance [23:36].
2. Local Caching Dramatically Lowers WAN Strain: Implementing Lancache eliminates redundant large file downloads, safeguarding latency and bandwidth for active gaming and streaming [26:58].
3. Next-Gen Goal (100G in a Small Box): The team aims to replace heavy 2U server chassis with ultra-compact hardware (e.g., DGX Spark / edge systems with dual 100G interfaces or GPU-accelerated DPDK packet processing) to handle 100G NAT and caching for future venue expansions at Taipei World Trade Center [27:41].

《COSCUP 2026 官網》https://coscup.org/2026