For the record: zabra's performance fix didn't have that much effect when the CPU was completely idle, but when loaded (hence, I presume, workers finishing at different times?) it moved my back-of-envelope PerformanceTest.test_performance timings from 58% efficiency (standard deviation of 8%) to 73% efficiency (standard deviation of 6%). That's using 8 samples on both, loading with some other performance test tasks. I'm pretty sure it's good.
For the record: zabra's performance fix didn't have that much effect when the CPU was completely idle, but when loaded (hence, I presume, workers finishing at different times?) it moved my back-of-envelope PerformanceTest.test_performance timings from 58% efficiency (standard deviation of 8%) to 73% efficiency (standard deviation of 6%). That's using 8 samples on both, loading with some other performance test tasks. I'm pretty sure it's good.