9% performance loss with 1.16.3 vs. 1.15.11

[Robert Cowart]

Having read about some of the enhancements in 1.16 we were expecting perhaps a minor performance increase from 1.16. However testing has revealed performance degradation of 9% compared to 1.15. We haven't dug into this deeper yet, but I was curious if anyone else has experienced anything similar?

The following shows the throughput in records per second with different Golang versions. For context, this is decoding and processing of network flow records (think of it as a simplified variant of protobuf), with a fixed number of go routines, and the same repeating dataset for each test. Compiling with 1.16.3 results in reduction in throughput of 5000 records/s (-9%) compared to 1.15.7 and 1.15.11.

Before digging into this further, I thought I would ask if anyone else has seen this?

Rob

[Brian Candler]

Could you compare with go 1.15 running with GODEBUG=madvdontneed=1 ?  That's now standard behaviour in 1.16.

[Robert Cowart]

I read up on that and agree that it seemed like a probable cause. However tests show no difference.