Benchmarking on Redis gets low performance when connection number raise to merely 5000

Question

Environment:

• Redis on a single machine (standalone mode) with 512GB mem and 128 cores.

Benchmark procedure:

1. run redis-benchmark -h xx -p xx -c 5000 -n 1000000 -t set,get, the result is like:
   • 
2. run redis-benchmark -h xx -p xx -c 1700 -n 1000000 -t set,get 3 times on the same server (split 5000 conns to 3 processes for executing), and the result on avg is roughly like:
   • 
3. run redis-benchmark -h xx -p xx -c 1700 -n 1000000 -t set,get only once, and the result is:
   • 

I've tried adding -P pipeline config, and it makes no big difference when comparing to the above result. I'm wondering why it suffers from performance penalty when configuring 5000 conns in a single redis-benchmark process? And how could I benchmark the real capability of current Redis instance? Thanks!