Over the past year or so various people have been automating kernel builds with the aim of both setting the standard that things should build reliably and using the resulting builds for automated testing. This has been having good results, it’s especially nice to compare the results for older stable kernel builds with current ones and notice how much happier everything is.
One of the challenges with doing this is that for good coverage you really need to include allmodconfig or allyesconfig builds to ensure coverage of as much kernel code as possible but that’s fairly resource intensive given the size of the kernel, especially when you want to cover several architectures. It’s also fairly important to get prompt results, development trees are changing all the time and the longer the gap between a problem appearing and it being identified the more likely the report is to be redundant.
Since I was looking at my own setup and I know of several people who’ve done similar benchmarking I thought I’d publish some ballpark numbers for from scratch allmodconfig builds on a single architecture:
|i7-4770 with SSD||20 minutes|
|linode 2048||1.25 hours|
|EC2 m3.medium||1.5 hours|
|EC2 c3.large||2 hours|
|Cubietruck with SSD||20 hours|
|Intel Celeron N2940||1.75 hours|
All with the number of tasks spawned by make set to the number of execution threads the system has and no speedups from anything like ccache. I may keep this updated in future with further results.
Obviously there’s tradeoffs beyond the time, especially for someone like me doing this at home with their own resources – my desktop is substantially faster than anything else I’ve tried but I’m also using it interactively for my work, it’s not easily accessible when not at home and the fans spin up during builds while EC2 starts to cost noticeable money to use as you add more builds.