Also there is some information about how to configure Linux environment for performance written in Brendan Gregg’s book “Systems Performance: Enterprise and the Cloud”, see chapter 6.8 CPU Tuning. Other very informal and inspiring articles on the subject: “Microbenchmarking calls for idealized conditions” and “Benchmarking tips”. I tried to sort the configuration settings in order of the impact on performance (according to my experience). Alternatively, advices in this article might be helpful if you want to compare performance of two snippets of code in a robust way. For example, if you’re a developer of performance critical application and you want to check if your change to the source code really did something good in terms of performance, then this article is exactly what you are looking for.
If you want to get absolute numbers to understand how your app will behave in the field, you should not make any artificial tuning to the system, as the client might have default settings.Īll the info in the article is applicable if you do the measurements on the same system. If you want to compare two different versions of the same program you should use suggestions described above. If you use all the advices in this article it is not how your application will run in practice. It is important that you understand one thing before we start. This article is an attempt to bring all the tips together, provide examples and give instructions how to configure your machine properly. So, ideally when doing benchmarking we try to disable all the potential sources of performance non-determinism in a system. That’s where variation in results might come from. However, the second run will operate on base frequency without entering the turbo mode. CPU will increase its frequency to the extent permitted by thermal package ( TDP) and then go back somewhere around its base frequency. Take turbo boost feature, for example: if we start two runs, one right after another on a “cold” processor, first run will possibly work for some time in overclocked mode.
Since we have little control over them, it makes sense to disable them to receive more consistent measurements and reduce the noise. In fact, today we will only talk about the features which are not permanently active. But some of them have non-deterministic behavior. Lots of features in HW and SW are intended to increase performance. Subscribe to my newsletter, support me on Patreon or by PayPal donation. 8) Use statistical methods to process measurements.3) Set scaling_governor to ‘performance’.
0 kommentar(er)
