https://pantsbuild.org/ logo
Powered by
#general
  • h

    happy-kitchen-89482

    09/24/2015, 11:43 PM
    of that, 25% (of total running time) is in abc.new, called in target_addressable:factory
  • h

    happy-kitchen-89482

    09/24/2015, 11:43 PM
    So all our proxies and callbacks etc are adding up
  • h

    happy-kitchen-89482

    09/24/2015, 11:43 PM
    there are hundreds of thousands of objects involved.
  • b

    bored-art-40741

    09/24/2015, 11:43 PM
    Proxy proliferation!
  • b

    bored-art-40741

    09/24/2015, 11:44 PM
    You can mostly thank me for this
  • h

    happy-kitchen-89482

    09/24/2015, 11:44 PM
    Thank you simple smile
  • b

    bored-art-40741

    09/24/2015, 11:44 PM
    ha
  • h

    happy-kitchen-89482

    09/24/2015, 11:44 PM
    I'm just trying to prove the need for an alternate BUILD file format and corresponding alternate build graph representation.
  • h

    happy-kitchen-89482

    09/24/2015, 11:44 PM
    Something much much much lighter weight.
  • m

    many-doctor-21484

    09/24/2015, 11:44 PM
    @happy-kitchen-89482: what's your choice of tools for profiling pants?
  • b

    bored-art-40741

    09/24/2015, 11:45 PM
    Oh hey @happy-kitchen-89482 you know what would be amazing
  • b

    bored-art-40741

    09/24/2015, 11:45 PM
    You can totally profile subprocesses
  • h

    happy-kitchen-89482

    09/24/2015, 11:45 PM
    @many-doctor-21484
  • h

    happy-kitchen-89482

    09/24/2015, 11:45 PM
  • b

    bored-art-40741

    09/24/2015, 11:45 PM
    If you set the right processing startup hook when the pool is started
  • b

    bored-art-40741

    09/24/2015, 11:46 PM
    Would have been handy when we were scratching our heads over the artifact bloat
  • b

    bored-art-40741

    09/24/2015, 11:46 PM
    It would also be nice if we started doing some sort of statmon equivalent thing in pants
  • h

    happy-kitchen-89482

    09/24/2015, 11:46 PM
    yeah
  • h

    happy-kitchen-89482

    09/24/2015, 11:46 PM
    gotta run, be back later
  • b

    bored-art-40741

    09/24/2015, 11:46 PM
    Count all the things
  • r

    rough-minister-58256

    09/25/2015, 12:05 AM
    cProfile doesn’t profile threads tho w/o extra work - this is my profiling shim using yappi: https://gist.github.com/kwlzn/42b809558dc825821ddb
  • r

    rough-minister-58256

    09/25/2015, 12:06 AM
    this also has the added benefit of callgrind output which has a number of visualization tools (kcachegrind/qcachegrind)
  • r

    rough-minister-58256

    09/25/2015, 12:06 AM
    @happy-kitchen-89482: ^
  • b

    bored-art-40741

    09/25/2015, 12:08 AM
    It would be cool to have that upstream too, then
  • r

    rough-minister-58256

    09/25/2015, 12:09 AM
    I also wrote this tracer thing too which may be useful for zeroing in on specific bits of code: https://github.com/kwlzn/pytracing
  • r

    rough-minister-58256

    09/25/2015, 12:10 AM
    we used it + some derivatives for analyzing the jvm compile scheduling optimizations
  • b

    bored-art-40741

    09/25/2015, 12:12 AM
    Hey speaking of the jvm compile scheduling!
  • b

    bored-art-40741

    09/25/2015, 12:12 AM
    One of our devs had a cool idea: we'd like to have a scheduling strategy where we "shuffle" the order that targets are put into queues, to whatever extent is allowed by the topological ordering
  • b

    bored-art-40741

    09/25/2015, 12:13 AM
    With cache double-checking this can improve parallelization of CI workers that all kick off at the same time
  • b

    bored-art-40741

    09/25/2015, 12:13 AM
    And are compiling overlapping subgraphs
1...461462463...2420Latest