Robin Shen [02/Apr/14 12:17 AM] This is a follow-up to http://track.pmease.com/browse/QB-1862 addressing significant wait time in transitioning between steps in the build flow. With a grid of 544 nodes we are seeing a correlation between Queue size, running builds and wait duration in a build with 3 simple steps printing "hello world", all configured to run on the same node as "master" step: builds in queue | builds running | build wait duration (s) 253 154 6,8 500 187 14,8 752 364 25,3 1021 415 32,2 1230 546 44,8 When we build 25000 times per day and the average number of steps is 10 and the average waiting time is 10 seconds, we spend 2.5 million seconds waiting every day. If every build is 10 minutes on the average and we spend 10*10 seconds idling, we would free up an equivalent of 500 build machines * 10*10 seconds / 10 minutes = 80+ build machines if the delay was reduced to zero.

Andrei Kortsak [02/Apr/14 06:56 AM] The most critical place for improvement for us would be transition time between steps, where "Node selection" = "On the same node running parent step".