Attention:

The measurements shown in the slides have incomplete stack traces, i.e., they show the wrong functions. Re-generating them didn't show bottlenecks.


Comments to slides:

• Breakdown of performance after 16 cores is likely caused by missing CPU pinning. Everything <= 16 is pinned to a single CPU. Afterwards, threads are free to move.
• Analysis is clearly I/O limited when coming from EOS. In an analysis that does more data processing, this would be less pronounced.
• Slide 17
  • Need to get rid of locks when sending requests to the type system.
  • Other lock seems to be waiting for EOS.
• Slide 18
  • Filter are being run bottom up instead of top down. Inverting would lead to huge performance gains in analyses that don't select many events.
  • This might become important because people will intuitively put restrictive filters first.