Performance Tuning (5,000+ Channels)

Without hyper-specific routing limits and meticulous micro-adjustments to the underlying environments, absorbing massive concurrent traffic spikes (5000+ simultaneous UDP streams) pushes servers into catastrophic TCP socket starvation, File Handle exhaustion, and ultimate cascading grid death blocks. This guide documents the exact kernel tuning parameters, ring-buffer allocations, and pipeline configuration structures necessary to smoothly offload extreme density network traffic, allowing systems to ingest tsunami-scale events without flinching.

Linux Kernel Network Stack Optimization

By default, standard Linux distributions are not optimized to handle the immense packet-per-second (pps) velocity generated by 5,000+ concurrent VoIP streams. You must bypass the default kernel network bottlenecks via sysctl tuning.

# Add to /etc/sysctl.conf for extreme UDP loads
net.core.rmem_max = 26214400      # Max receive buffer (25MB)
net.core.rmem_default = 26214400 
net.core.netdev_max_backlog = 100000 # Queue size for incoming packets
net.ipv4.udp_mem = 65536 131072 262144 # UDP memory boundaries
net.ipv4.udp_rmem_min = 16384     # Minimum read buffer per socket

Ingestion Engine Go-Routine Scaling

Inside /etc/cxmind/config.yaml, you must dramatically expand the default concurrent processing limits. The DSP pools decode G.711/Opus payloads natively via highly optimized Go routines. Starving these pools leads to immediate audio clipping and catastrophic ASR transcription failure.

• dsp_thread_pool: Increase from default `200` to `20000`. Memory footprint per routine is roughly ~2KB, making high numbers completely safe and strictly CPU-bound.
• asr_connection_multiplex: Each upstream ASR WebSocket can multiplex audio. Set this limit matching upstream vendor ratelimits (default `10` streams per WebSocket tunnel).
• redis_pipeline_batch_size: Crucial for preventing massive Redis latency spikes. Set to `500` to enforce highly compressed multi-exec Pub/Sub broadcasts instead of spamming rapid single events.