Garbage Collection
Joydeep Dey
slide 1
Major Areas of Memory
Static area
Fixed size, fixed content, allocated at compile time
Run-time stack
Variable size, variable content (activation records) Used for managing function calls and returns
Heap
Fixed size, variable content Dynamically allocated objects and data structures
Examples: ML reference cells, malloc in C, new in Java
slide 2
Cells and Liveness
Cell = data item in the heap
Cells are pointed to by pointers held in registers, stack, global/static memory, or in other heap cells
Roots: registers, stack locations, global/static variables A cell is live if its address is held in a root or held by another live cell in the heap
slide 3
Garbage
Garbage is a block of heap memory that cannot be accessed by the program
An allocated block of heap memory does not have a reference to it (cell is no longer live) Another kind of memory error: a reference exists to a block of memory that is no longer allocated
Garbage collection (GC) - automatic management of dynamically allocated storage
Reclaim unused heap blocks for later use by program
slide 4
Example of Garbage
class node { int value; node next; } node p, q; p = new node(); q = new node(); q = p; delete p;
slide 5
Why Garbage Collection?
Todays programs consume storage freely
1GB laptops, 1-4GB deskops, 8-512GB servers 64-bit address spaces (SPARC, Itanium, Opteron)
and mismanage it
Memory leaks, dangling references, double free, misaligned addresses, null pointer dereference, heap fragmentation Poor use of reference locality, resulting in high cache miss rates and/or excessive demand paging
Explicit memory management breaks high-level programming abstraction
slide 6
GC and Programming Languages
GC is not a language feature GC is a pragmatic concern for automatic and efficient heap management
Cooperative langs: Lisp, Scheme, Prolog, Smalltalk Uncooperative languages: C and C++
But garbage collection libraries have been built for C/C++
Recent GC revival
Object-oriented languages: Modula-3, Java
In Java, runs as a low-priority thread; System.gc may be called by the program
Functional languages: ML and Haskell
slide 7
The Perfect Garbage Collector
No visible impact on program execution Works with any program and its data structures
For example, handles cyclic data structures
Collects garbage (and only garbage) cells quickly
Incremental; can meet real-time constraints
Has excellent spatial locality of reference
No excessive paging, no negative cache effects
Manages the heap efficiently
Always satisfies an allocation request and does not fragment
slide 8
Summary of GC Techniques
Reference counting
Directly keeps track of live cells GC takes place whenever heap block is allocated Doesnt detect all garbage
Tracing
GC takes place and identifies live cells when a request for memory fails Mark-sweep Copy collection
Modern techniques: generational GC
slide 9
Reference Counting
Simply count the number of references to a cell Requires space and time overhead to store the count and increment (decrement) each time a reference is added (removed)
Reference counts are maintained in real-time, so no stop-and-gag effect Incremental garbage collection
Unix file system uses a reference count for files C++ smart pointer (e.g., auto_ptr) use reference counts
slide 10
Reference Counting: Example
Heap space root set
1 2
slide 11
Reference Counting: Strengths
Incremental overhead
Cell management interleaved with program execution Good for interactive or real-time computation
Relatively easy to implement Can coexist with manual memory management Spatial locality of reference is good
Access pattern to virtual memory pages no worse than the program, so no excessive paging
Can re-use freed cells immediately
If RC == 0, put back onto the free list
slide 12
Reference Counting: Weaknesses
Space overhead
1 word for the count, 1 for an indirect pointer
Time overhead
Updating a pointer to point to a new cell requires:
Check to ensure that it is not a self-reference Decrement the count on the old cell, possibly deleting it Update the pointer with the address of the new cell Increment the count on the new cell
One missed increment/decrement results in a dangling pointer / memory leak Cyclic data structures may cause leaks
slide 13
Reference Counting: Cycles
Heap space root set
Memory leak
1 1
slide 14
Smart Pointer in C++
Similar to std::auto_ptr<T> in ANSI C++
Ref<T> x RefObj<T> *ref RefObj<T> T* obj: int cnt: 2 object of type T
RefObj<T> *ref
sizeof(RefObj<T>) = 8 bytes of overhead per reference-counted object
sizeof(Ref<T>) = 4 bytes Fits in a register Easily passed by value as an argument or result of a function Takes no more space than regular pointer, but much safer (why?)
slide 15
Smart Pointer Implementation
template<class T> class Ref { template<class T> class RefObj { RefObj<T>* ref; T* obj; Ref<T>* operator&() {} int cnt; public: public: Ref() : ref(0) {} RefObj(T* t) : obj(t), cnt(0) {} Ref(T* p) : ref(new RefObj<T>(p)) { ref->inc();} ~RefObj() { delete obj; } Ref(const Ref<T>& r) : ref(r.ref) { ref->inc(); } ~Ref() { if (ref->dec() == 0) delete ref; } int inc() { return ++cnt; } int dec() { return --cnt; } Ref<T>& operator=(const Ref<T>& that) { if (this != &that) { if (ref->dec() == 0) delete ref; ref = that.ref; ref->inc(); } return *this; } T* operator->() { return *ref; } T& operator*() { return *ref; } };
slide 16
operator T*() { return obj; } operator T&() { return *obj; } T& operator *() { return *obj; } };
Using Smart Pointers
Ref<string> proc() { Ref<string> s = new string(Hello, world); // ref count set to 1 int x = s->length(); // s.operator->() returns string object ptr return s; } // ref count goes to 2 on copy out, then 1 when s is auto-destructed int main() { Ref<string> a = proc(); // ref count is 1 again } // ref count goes to zero and string is destructed, along with Ref and RefObj objects
slide 17
Mark-Sweep Garbage Collection
Each cell has a mark bit Garbage remains unreachable and undetected until heap is used up; then GC goes to work, while program execution is suspended Marking phase
Starting from the roots, set the mark bit on all live cells
Sweep phase
Return all unmarked cells to the free list Reset the mark bit on all marked cells
slide 18
Mark-Sweep Example (1)
Heap space root set
slide 19
Mark-Sweep Example (2)
Heap space root set
slide 20
Mark-Sweep Example (3)
Heap space root set
slide 21
Mark-Sweep Example (4)
Heap space root set
Free unmarked cells
Reset mark bit of marked cells
slide 22
Mark-Sweep Costs and Benefits
Good: handles cycles correctly Good: no space overhead
1 bit used for marking cells may limit max values that can be stored in a cell (e.g., for integer cells)
Bad: normal execution must be suspended Bad: may touch all virtual memory pages
May lead to excessive paging if the working-set size is small and the heap is not all in physical memory
Bad: heap may fragment
Cache misses, page thrashing; more complex allocation
slide 23
Copying Collector
Divide the heap into from-space and to-space Cells in from-space are traced and live cells are copied (scavenged) into to-space
To keep data structures linked, must update pointers for roots and cells that point into from-space
This is why references in Java and other languages are not pointers, but indirect abstractions for pointers
Only garbage is left in from-space
When to-space fills up, the roles flip
Old to-space becomes from-space, and vice versa
slide 24
Copying a Linked List
from-space
root A
[Cheneys algorithm]
pointer forwarding address
C
D
to-space
A B C D
Cells in to-space are packed
slide 25
Flipping Spaces
to-space
pointer forwarding address
from-space
root A B C D
slide 26
Copying Collector Tradeoffs
Good: very low cell allocation overhead
Out-of-space check requires just an addr comparison Can efficiently allocate variable-sized cells
Good: compacting
Eliminates fragmentation, good locality of reference
Bad: twice the memory footprint
Probably Ok for 64-bit architectures (except for paging)
When copying, pages of both spaces need to be swapped in. For programs with large memory footprints, this could lead to lots of page faults for very little garbage collected Large physical memory helps
slide 27
Generational Garbage Collection
Observation: most cells that die, die young
Nested scopes are entered and exited more frequently, so temporary objects in a nested scope are born and die close together in time Inner expressions in Scheme are younger than outer expressions, so they become garbage sooner
Divide the heap into generations, and GC the younger cells more frequently
Dont have to trace all cells during a GC cycle Periodically reap the older generations Amortize the cost across generations
slide 28
Generational Observations
Can measure youth by time or by growth rate Common Lisp: 50-90% of objects die before they are 10KB old Glasgow Haskell: 75-95% die within 10KB
No more than 5% survive beyond 1MB
Standard ML of NJ reclaims over 98% of objects of any given generation during a collection C: one study showed that over 1/2 of the heap was garbage within 10KB and less than 10% lived for longer than 32KB
slide 29
Example with Immediate Aging (1)
C root set
A
B
Young
F G
Old
slide 30
Example with Immediate Aging (2)
C root set E
Young
F B
Old
slide 31
Generations with Semi-Spaces
root set
Youngest . . . From-space To-space
Middle generation(s)
Oldest From-space To-space
slide 32
