Substrate state cache and performance limitation?

Question

I've been testing Substrate with my own configuration (latest node-template v3 fork, with an additional basic pallet) and trying to understand a drop in performances.

In this screenshot 2, you can see 5 times the exact same test (sending ~ 200tx/sec). I've tested multiple times and each time it looks like, when hitting a state cache of 20MB, the total transaction finished rate performances decrease and aren't stable at all like the first test.

• What is happening? Why 5 times the same tests we don't get the same result?
• Is there a link between state cache and performances?
• If increasing --state-cache-size increases performances?

My node command line looks like:

node-template \
--base-path /datas/substrate-0 \
--name Node0 \
--chain /genesis/customSpecRaw.json \
--port 30333 \
--ws-port 9944 \
--rpc-port 9933 \
--node-key 0000000000000000000000000000000000000000000000000000000000000001 \
--unsafe-ws-external \
--unsafe-rpc-external \
--rpc-cors=all \
--prometheus-external \
--log info \
--wasm-execution Compiled \
--ws-max-connections 1000 \
--pool-limit 10000 \
--pool-kbytes 20480 \
--validator \
--max-runtime-instances 100

Screenshot: