Hmm, it seems though you're right, it seemed odd to me that the difference between puma would be so large at high concurrency, but when I look at it now the I see how the test does correctly mimic I/O for the purpose of the benchmark.

My apologies if I came out too strong.