Issue #285, Fix rate-limiting with cluster-mode

[YaacovHazan]

When a connection disconnected, the timer event was not free, and cause the test to keep running forever.

One of these cases is when we are starting the benchmark in cluster-mode and using the replica's ip/port.

[LINKIWI]

Hi, we believe that this commit may have severely regressed cluster mode performance in memtier_benchmark. After upgrading from 2.1.1 to 2.1.3, our performance tests are indicating a nearly 50% drop in throughput and 10x higher P99.9. This PR appears to be associated with the only commit in 2.1.3 that meaningfully changes the benchmark's core logic.

[filipecosta90]

@LINKIWI we have identified #294 as the root cause, and have a PR to fix it #295.

Altogether, that will revert the regression introduced in v2.1.1 and also reduced by 22% the CPU usage for the same QPS (even a bit higher). Can you check if that PR will fix your regression while we're reviewing it?

[LINKIWI]

Confirming that #295 applied on top of 2.1.3 indeed fixes the perf regression we are observing. Thanks! (Apologies for jumping to conclusions on the root cause; we didn't have time yesterday to dive deeper after noticing the original issue.)