[QCon 2019] The Trouble With Memory

Kirk Pepperdine

For other QCon blog posts, see QCon live blog table of contents

General

  • Slow database queries, inefficient app code and too many database queries are most reported problems
  • Once drill down, over 70% of all Java apps are bottlenecked on memory churn. It’s not reported because hard to observe
  • Tend to put logging around past problems.
  • If apply instrument to a system, it will always tell you something. And then you act on it
  • Cheapar to predict than react

Common libraries

  • Logback
  • Marshalling Json, SQL
  • Caching products
  • Hibernate

Memory

  • Java heap has generations
  • Hopefully people have moved to G1GC
  • Everything happens in the free list

Problems

  • Large number of temporary objects quickly fills Eden
  • Causes frequent young cycles. Causes premature promotion which means will go to tenured too early
  • Heap becomes more fragment
  • Allocation is quick. No cost to collect if objects die quickly. However, still slow if you do something quick enough times.
  • Large live data set size. Data consistently live in your heap. Increases time to copy/compact. Likely have less space to copy to. Think about Windows defragmenter. [Do people still have to do that?]
  • Memory leak from unstable live data. JVM will terminate if you are lucky.
  • Out of memory – 98% of recent time spent in GC with less than 2% of heap recovered. If don’t meet that criteria, app is just really slow, but don’t get the out of memory error.

Escape analysis

  • Test applied to a piece of data. What is the visibility/scope.
  • If scoped locally, only thread that created it can see it.
  • If passed to method, partial escape.
  • If data scoped so multiple threads can see it (Ex: static), full escape.

Demo

  • Showed GC log. Want to see low pause times
  • Showed allocation rates. Problem if too high
  • In Visual VM, looked at profiler. Check filters to ensure not filtering the bottleneck out of your profile
  • Sort by # allocated objects to see frequency. It doesn’t take longer to allocate a large object than a small one.
  • Take a snapshot and look at trace
  • “Stop thinking” – explore what is shown without assuming
  • Time to look at the code from the stack trace that is creating all the objects
  • Escape analysis code
  • Run jitwatch to see allocations. Can see if direct/inline allocation. Can see when bytecode eliminates an allocation
  • Profiler is lying to you.
  • Performance differs in test vs prod environment

Q&A

  • How know the performance problem is the int[] in the demo? Went through profiler to show stack trace. Used BigInteger which uses up a lot more memory than a long
  • Absolute number for GC allocation rate? Sparc? Number seem to hold regardless of hardware. Should focus on the CPU going forward.
  • <missed question> – try to find mutable state that is not shared

My impression

This was great. I learned a lot and it kept my attention. I really liked the demo.

Leave a Reply

Your email address will not be published.