use threadlocal random generator to avoid contention

[eivanov89]

Hi,

I noticed a very high CPU consumption, when import initial TPC-C data. According the attached flamegraph all threads are in TPCCUtil:::randomStr. Because of util.RandomGenerator there is a contention. This patch uses thread local random generator, which helps to avoid contention and dramatically reduces CPU consumption.

[apavlo]

@eivanov89 Nice! I will take a look

[eivanov89]

@apavlo
Attached missing flamegraph.

[apavlo]

I am trying to decide if we should just drop com.oltpbenchmark.util.RandomGenerator and replace it with ThreadLocalRandomGenerator. We can just replace that implementation with the new one (using the same name).

I don't think there will be an non-determinism issue...

[eivanov89]

@apavlo it seems that we need your approval to run the workflow.

[eivanov89]

@apavlo ping

[apavlo]

@eivanov89 Thanks for poking me. I will look at this in more detail later this week.

[apavlo]

@eivanov89 Let's have your RNG class replace the existing one. Can you do this?

[eivanov89]

@eivanov89 Let's have your RNG class replace the existing one. Can you do this?

@apavlo Sure, I can try to do it next week.

[eivanov89]

@eivanov89 Let's have your RNG class replace the existing one. Can you do this?

@apavlo I'm sorry for a long delay. Here is the updated pull request.