Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems G. Salagnac, C. Rippert, S. Yovine Verimag Lab. Université Joseph Fourier Grenoble, France http://www-verimag.imag.fr/~salagnac [email protected] G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 1 / 16 Motivation The Java programming language I Attractive language I Automatic memory management Implementation pitfalls I Garbage Collection ⇒ pause times and fragmentation =⇒ Difficult to use in a real-time embedded context G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 2 / 16 Our approach Non-determinism of Garbage Collector pause times: the problem is in the JVM, not in the language! Proposition: I Keep the language I no manual memory management I Change the implementation I replace the GC by a predictable allocator I use region-based memory management I automatically compute object lifetimes at compilation I undecidable problem I find a reasonable over-approximation G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 3 / 16 Our approach Non-determinism of Garbage Collector pause times: the problem is in the JVM, not in the language! Proposition: I Keep the language I no manual memory management I Change the implementation I replace the GC by a predictable allocator I use region-based memory management I automatically compute object lifetimes at compilation I undecidable problem I find a reasonable over-approximation G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 3 / 16 Outline Introduction Region-Based Memory management Pointer Interference Analysis Experimental results Conclusion G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 4 / 16 Memory management using regions a java heap... ...organised in regions region 1 region 2 program variables Objects in a region will share the same (physical) lifetime I Benefits: a more real-time-compatible behaviour I objects allocated side by side I no fragmentation, constant time I region destroyed as a whole: predictable times I Drawbacks: more difficult bookkeeping I object placement issue: who decides? I region destroyed as a whole: space overhead, risk of faults G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 5 / 16 Memory management using regions a java heap... ...organised in regions region 1 region 2 program variables Objects in a region will share the same (physical) lifetime I Benefits: a more real-time-compatible behaviour I objects allocated side by side I no fragmentation, constant time I region destroyed as a whole: predictable times I Drawbacks: more difficult bookkeeping I object placement issue: who decides? I region destroyed as a whole: space overhead, risk of faults G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 5 / 16 Memory management using regions a java heap... ...organised in regions region 1 region 2 program variables Objects in a region will share the same (physical) lifetime I Benefits: a more real-time-compatible behaviour I objects allocated side by side I no fragmentation, constant time I region destroyed as a whole: predictable times I Drawbacks: more difficult bookkeeping I object placement issue: who decides? I region destroyed as a whole: space overhead, risk of faults G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 5 / 16 Outline Introduction Region-Based Memory management Pointer Interference Analysis Experimental results Conclusion G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 6 / 16 Region analysis and allocation policy Hypothesis: Objects within the same data structure (i.e. connected together) will often have similar (logical) lifetimes =⇒ one region for each data structure I no inter-region pointer Static analysis: I identify variables that may point to connected objects Allocation Policy: I place objects so that each structure is grouped in a region G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 7 / 16 Region analysis and allocation policy Hypothesis: Objects within the same data structure (i.e. connected together) will often have similar (logical) lifetimes =⇒ one region for each data structure I no inter-region pointer Static analysis: I identify variables that may point to connected objects Allocation Policy: I place objects so that each structure is grouped in a region G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 7 / 16 ∼ ∼ ∼ ∼ list o1 o2 this tmp this o main() <init>() add() Example: pointer interference analysis class ArrayList { Object[] data; int index; main() { <init>(int capacity) ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } Object o2=new Object; void add(Object o) list.add(o1); { } this.data[this.index] = o; this.index ++; } } G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ ∼ ∼ ∼ list o1 o2 this tmp this o Example: pointer interference analysis for all methods, class ArrayList { Object[] data; int index; main() main() { <init>(int capacity) ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ ∼ ∼ ∼ this tmp this o Example: pointer interference analysis identify local variables class ArrayList { Object[] data; int index; main() main() { <init>(int capacity) list o1 o2 ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ ∼ ∼ ∼ this o Example: pointer interference analysis identify local variables class ArrayList { Object[] data; int index; main() main() { <init>(int capacity) list o1 o2 ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) this tmp list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ ∼ ∼ ∼ Example: pointer interference analysis identify local variables class ArrayList { Object[] data; int index; main() main() { <init>(int capacity) list o1 o2 ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) this tmp list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() this o G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ ∼ ∼ Example: pointer interference analysis local interference: class ArrayList v1.f=v2 =⇒ v1∼v2 { Object[] data; int index; main() main() { <init>(int capacity) list o1 o2 ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) this ∼ tmp list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() this o G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ ∼ Example: pointer interference analysis local interference: class ArrayList v1.f=v2 =⇒ v1∼v2 { Object[] data; int index; main() main() { <init>(int capacity) list o1 o2 ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) this ∼ tmp list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() this ∼ o G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ Example: pointer interference analysis interproc. interference: class ArrayList p1∼p2 =⇒ a1∼a2 { Object[] data; int index; main() main() { <init>(int capacity) list ∼ o1 o2 ArrayList list = { new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) this ∼ tmp list.add(o1); { } this.data[this.index] = o; this.index ++; } } add() this ∼ o G. Salagnac (Verimag) Semi-Automatic Region-Based Memory Management for Real-Time Java Embedded Systems 8 / 16 ∼ Example: pointer interference analysis results: class ArrayList { Object[] data; int index; main() main() {//list∼o1 o2 <init>(int capacity) list ∼ o1 o2 ArrayList list = {//this∼tmp new ArrayList; this.index = 0; list.<init>(3); tmp = new Object[capacity]; this.data = tmp; Object o1=new Object; } <init>() Object o2=new Object; void add(Object o) this ∼ tmp