Runtime: Microbenchmarks should run with tiered JITing enabled
From https://github.com/dotnet/coreclr/issues/22128#issuecomment-456513621
Microbenchmarks runs are done with tiered JITing disabled. It means that they are not measuring performance that our customers are going to see in the default configuration.
We should have microbenchmarks run with the default configuration (ie tiered JITing enabled).
jkotas
All 4 comments
cc @jorive
jkotas
on 22 Jan 2019
@billwert @adamsitnik Is this already covered in dotnet/performance? If not, will it be?
RussKeldorph
on 14 Jun 2019
In the old perf lab, the tiered JIT was always disabled.
At the end of February, we have added runs for tiered and non-tiered micro benchmark runs to the performance repo https://github.com/dotnet/performance/pull/340 Initially, the gap was huge for some of the benchmarks with "hot loops"
After @kouvel has merged https://github.com/dotnet/coreclr/pull/23599 the gap dissapeared.
The current status is that we are replacing the "Tiered" and "non-Tiered" runs with single run with default settings: https://github.com/dotnet/performance/pull/561
adamsitnik
on 14 Jun 2019
@adamsitnik Thanks!
RussKeldorph
on 15 Jun 2019
Related issues
omajid
路
3Comments
matty-hall
路
3Comments
omariom
路
3Comments
nalywa
路
3Comments
Timovzl
路
3Comments
Most helpful comment
In the old perf lab, the tiered JIT was always disabled.
At the end of February, we have added runs for tiered and non-tiered micro benchmark runs to the performance repo https://github.com/dotnet/performance/pull/340 Initially, the gap was huge for some of the benchmarks with "hot loops"
After @kouvel has merged https://github.com/dotnet/coreclr/pull/23599 the gap dissapeared.
The current status is that we are replacing the "Tiered" and "non-Tiered" runs with single run with default settings: https://github.com/dotnet/performance/pull/561