Hide Notes 

Slide 1: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen SGen — Mono’s new Garbage Collector Mark Probst Novell / LINBIT 04 February 2011
Slide 2: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Outline 1 Introduction 2 Simple garbage collection 3 Generational garbage collection 4 SGen 5 Improving your program’s performance 6 Tuning SGen
Slide 3: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Me On the Mono team since 2007 Generic code sharing Initial PPC64 port Working on SGen since April 2009
Slide 4: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Why not Boehm for Life? Slow allocation Not generational Hard to do precise GC with App Domains Fragmentation issues
Slide 5: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 6: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 7: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 8: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 9: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 10: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 11: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Root R
Slide 12: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 13: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 14: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 15: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 16: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 17: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 18: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Copying Root R
Slide 19: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen The generational hypothesis Most objects die young Those that don’t live for a long time
Slide 20: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Generations Separate heaps for young and old objects Young objects are collected frequently Those that survive get "promoted"
Slide 21: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Write barriers Nursery Root Major Heap
Slide 22: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Write barriers Nursery Root Major Heap
Slide 23: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Generations Nursery Major heap Large Object Space (LOS) For objects ≥ 8000 bytes Logically part of the major heap
Slide 24: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Thread Local Allocation Buffers (TLAB) Nursery Thread 1 Thread 2 Thread 3 Each thread gets a small part of the nursery Uncontested pointer-bumping allocation
Slide 25: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Pinning Not everything is scanned precisely Even then, not all precisely scanned pointers can be changed Everything on the stack
Slide 26: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Pinning We cannot move some objects This is problematic mostly in the nursery Because it’s small and fast allocation depends on it
Slide 27: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Mark-and-Sweep Fixed-size blocks (16kb) Each block contains objects of the same size That helps prevent fragmentation
Slide 28: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Evacuation
Slide 29: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Allocation is cheap p u b l i c c l a s s Test { int a , b ; p u b l i c s t a t i c i n t Main ( ) { f o r ( i n t i = 0 ; i < 5 0 0 0 0 0 0 0 ; ++i ) T e s t t = new T e s t ( ) ; return 0; } } Boehm SGen SGen 128k nursery Quad Core i7 Darwin 4s 11498 GCs (1.5s) 0.45s 191/1 GCs (3.6ms) 0.8s 6141/1 GCs (69ms) Dual Core i7 Linux 3s 17321 GCs (1.7s) 0.5s 287/1 GCs (6ms) 1.1s 9215/1 GCs (152ms)
Slide 30: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Don’t use free lists Live objects are a continuous cost Unless you don’t do any other allocation at all Even then, see previous slide
Slide 31: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Don’t write to reference fields Especially not in old objects Better yet, don’t have them in the first place Objects without reference fields are processed faster
Slide 32: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Weak refs are expensive But less expensive than on Boehm
Slide 33: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Finalization is expensive (and evil) Time of finalization is not determined Finalization order is not determined Objects are kept alive longer
Slide 34: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Environment Variable MONO_GC_PARAMS See man page
Slide 35: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Nursery Size nursery-size= must be a power of two Default is 4m Going too low leads to over-pinning Enlarge if you have more threads
Slide 36: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Major Collector major= marksweep The cautious default marksweep-par marksweep-fixed Use this unless you have good reason not to marksweep-fixed-par Better yet, use this and report bugs copying
Slide 37: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Concurrent Sweep no-concurrent-sweep Currently the default concurrent-sweep Use this, report bugs
Slide 38: SGen — Mono’s new Garbage Collector Mark Probst Introduction Simple garbage collection Generational garbage collection SGen Improving your program’s performance Tuning SGen Write Barrier wbarrier= cardtable The default Slow for large heaps remset Very bad if you do lots of reference field writes Good if you do very few Subject to short term change!