start distinct Go process per 2 (virtual) cores.

[funny-falcon]

It is very similar to usage nodejs cluster package or Ruby Unicorn server, so i think it is legitimate.

It is simple version, just to benchmark, but it could be easily extended to match Ruby Unicorn or nodejs cluster functionality.

I think it is able to greatly improve Go performance on Peak and EC2. Even i7 benchmark can benefit from this change.

[methane]

While I'm interested in effect of this change, I think preforking is not a usual way in Go's world.
I want to keep our code in good shape: new Gophers can use our code as sample code for writing production ready server.

I want to see future version of Go resolve scaling issue.
Go 1.4 has lock contention problem in database/sql. It is found in this benchmark and I've fixed it.
I'll try to investigate this problem before Go 1.6 code freeze, too.

[funny-falcon]

But "prefork" helps even with "plain text" and "json" benchmarks. So it is not tied to database/sql.

If i made it as a public library, will you use it for benchmark? Perhaps, as a separate "framework".

Or as external tool like https://github.com/stripe/einhorn with library as adapter for?

[methane]

First of all, this PR may not be merged Round 10 because Round 10 is in final phase for now.
(But I don't know, because I'm not a member of TechEmpower).
So we have enough time to chose better way.

General requirements says:

All test implementations should be production-grade. The particulars of this will vary by framework and platform, but the general sentiment is that the code and configuration should be suitable for a production deployment.

So I want our code and configuration are good enough for reference.
Using library or not is not a problem. The problem is: Is prefork a recommended / commonly used technique for high load production environment?

If the answer is yes, I'm OK to merge this.
If not, I think preforking benchmark should be separated from normal "go" benchmark. "go-prefork", maybe.

Split benchmark doesn't require split the code. Adding -prefork flag is enough, maybe.
I don't like einhorn since it depends Ruby. It makes hard to run test for new contributers.

[msmith-techempower]

I agree with @methane

[funny-falcon]

I've modified code, so by default it runs in one process.

If flag -prefork passed then prefork performed. It will be great, if it is added as go-prefork benchmark.

[hamiltont]

Split benchmark doesn't require split the code. Adding -prefork flag is enough, maybe.

If this is good with both of you, it's good with our team. You'll need to add the new configuration to the benchmark_config, and create a new setup file like setup_prefork to do exactly what setup.sh does but also pass the -prefork flag

[funny-falcon]

Ok, i've edit benchmark_config and add setup_prefork.sh . Hope i've done it right.

[hamiltont]

@funny-falcon Yup, that looks correct to me! We will now see what travis says

[funny-falcon]

looks like travis doesn't mind.

[funny-falcon]

:( why results could not be viewed?

[methane]

I suppose it because it's too late for Round 10.
Please wait for round 11.

FYI, Go 1.5 improved schedular:
golang/go@f47e581

So performance of Go without prefork will be improved slightly.