That controls the concurrency so it should slow down while reaching CPU/memory/event loop utilization. But spikes can still happen if you have bad luck.
First look at the logs, is concurrency rapidly increasing before you see the spike? We need more insight into what those spikes are