Reduce benchmark harness overhead

[arthurprs]

I was hitting some significant bottlenecks due to the benchmark harness overhead. I narrowed it down to the 2 limitations below, which lead to skewed results:

• overhead of parsing each line and generating a command
• single producer doing the command generation cannot keep the consumer threads busy

This change takes the easy (pragmatic) route and fully pre-process the benchmark commands before performing the benchmark. This solves both of the issues above.

Although effective, this change isn't particularly ideal, as it requires 4GB memory to benchmark the DS1 dataset.

[tatsuya6502]

Nice catch!

Although effective, this change isn't particularly ideal, as it requires 4GB memory to benchmark the DS1 dataset.

I am okay with that. Thank you for working on this enhancement.