Strange Latency while testing Pulsar

[wenze159632]

When I use the pulsar-perf tool to test the performance of Pulsar, I encounter the following strange phenomenon.

As shown in the image above, a small number of messages experience unacceptably high latency. Similar situations occur when testing in other clusters. My question is, what is the cause of this phenomenon and are there any areas that can be optimized?
The test commamd is

bin/pulsar-perf produce -n 1 -r 2000 -s 24576 -np 6 -m 500000  -bm 2000 test-t2

When the configuration '-bm' is removed, this high latency phenomenon occurs more frequently.

[wenze159632]

By setting -bm 2000, the frequency of occurrence of such high-latency messages can be reduced. If not set, the frequency of this problem occurring will greatly increase.
This phenomenon can also be viewed in Grafana panels.