A case study measuring throughput, CPU utilization, computational efficiency, and power consumption on a 32-core x64 server with a dual-port 100 GbE NIC. Using iperf3 with NUMA-aware pinning, single-core performance reached ~24 Gbps, while 8 parallel instances nearly saturated the 100 Gbps link. Scaling beyond 8 instances revealed a hidden NIC hardware bottleneck: an internal switch capped aggregate throughput at 100 Gbps across all ports, despite marketing materials omitting this limit. Power analysis shows CPU accounts for ~50% of total system draw, and system-level energy efficiency improves with higher parallelism even as per-core computational efficiency declines — a fundamental trade-off for server sizing decisions.
Sort: