On 5/27/26 09:23, Kurt Jaeger wrote: > Hi! > >>> Again, it's a dependency tree. You can not parallelize more >>> than a certain number, because nodes would have to wait for other nodes >>> to finish before they can work. So one has to reduce >>> the dependencies and shorten the longest path. > >> I do not see in void's wording anything indicating that the unit of work >> would be a builder run of one port-package-build. > > Yes. But if you have 10000 cores of the same GHz Rating as > one fast box with only, let's say, 128 cores. > > If you want to measure from the first to the last package, > the way to shorten the total build time is to analyse and > optimize the longest path in the dependency tree. > > . . .
FYI . . . One thing we do have evidence for are examples like (beefy23 building 150-amd64-default [so: latest] port-packages in both): ) an incremental build: built 1962 failed 86 elapsed 20:37:56 ) a from-scratch build: built 36937 failed 82 elapsed 29:49:46 So that about 6% in isolation took about 69% of the time of a from-scratch build. Specifically: Started 17 May 2026 01:01:18 GMT built 1962 failed 86 elapsed 20:37:56 Started 21 May 2026 01:01:20 GMT built 36937 failed 82 elapsed 29:49:46 No variations on numbers of cores/cpus or cpu/RAM clocking frequencies or such. No overheads of remote distribution and collection. And so on. Normally each port-package builder is limited to 3 or 4 make jobs at most (when the port respects such, some do not). There can be lots of idle core time when useful work could have been available and worked on. (But that gets into managing use of other resources, including to avoid failures from running out.) -- === Mark Millard marklmi at yahoo.com
