Throughput vs. client count

800 Mb/s per client looks fine in isolation, but at scale it strands core bandwidth and hides CPU/queue limits.

Quick take

• Right-size NICs and queue depth so aggregate throughput scales linearly with clients.
• Watch per-core IRQ load; rebalance or pin queues.
• Prefer larger I/O and fewer flows for backbone efficiency.

What to measure

Per-client goodput, tail latency (p99), NIC queue occupancy, drops, and CPU softirq time. Increase clients until p99 rises >20%—that’s your tipping point.