Simple bootstrap leads to perf regression compared to allocator-based bootstrap

[pzhan9]

When working on migrating allocator-based bootstrap to simple bootstrap for the benchmark (D93664956), we saw consistent regression in latency. Here are some experiments:

https://www.internalfb.com/servicelab/experiment/7001413966
https://www.internalfb.com/servicelab/experiment/6801405203
https://www.internalfb.com/servicelab/experiment/902416375977145

Here is a screenshot of one experiment: