Automatic Storage Management - PowerPoint PPT Presentation

About This Presentation
Title:

Automatic Storage Management

Description:

Automatic Storage Management. Patrick Earl. Simon Leonard. Jack Newton ... Baker's Read Barrier Algorithms. Incremental Copying. Non-copying Algorithm (The Treadmill) ... – PowerPoint PPT presentation

Number of Views:233
Avg rating:3.0/5.0
Slides: 103
Provided by: jack223
Category:

less

Transcript and Presenter's Notes

Title: Automatic Storage Management


1
Automatic Storage Management
  • Patrick Earl
  • Simon Leonard
  • Jack Newton

2
Overview
  • Terminology
  • Why use Automatic Storage Management?
  • Comparing garbage collection algorithms
  • The Classic algorithms
  • Copying garbage collection
  • Incremental Tracing garbage collection
  • Generational garbage collection
  • Conclusions

3
Terminology
  • Stack a memory area where activation records or
    frames are pushed onto when a procedure is called
    and popped off when it returns
  • Heap a memory area where data structures can be
    allocated and deallocated in any order.

4
Terminology(Continued)
  • Roots values that a program can manipulate
    directly (i.e. values held in registers, on the
    program stack, and global variables.)
  • Node/Cell/Object an individually allocated piece
    of data in the heap.
  • Children Nodes the list of pointers that a given
    node contains.
  • Live Node a node whose address is held in a root
    or is the child of a live node.

5
Terminology(Continued)
  • Garbage nodes that are not live, but are not
    free either.
  • Garbage collection the task of recovering
    (freeing) garbage nodes.
  • Mutator The program running alongside the
    garbage collection system.

6
Why Garbage Collect?
  • Language requirements
  • In some situations it may be impossible to know
    when a shared data structure is no longer in use.

7
Why Garbage Collect?(Continued)
  • Software Engineering
  • Garbage collection increases abstraction level of
    software development.
  • Simplified interfaces and decreases coupling of
    modules.
  • Studies have shown a significant amount of
    development time is spent on memory management
    bugs Rovner, 1985.

8
Comparing Garbage Collection Algorithms
  • Directly comparing garbage collection algorithms
    is difficult there are many factors to
    consider.
  • Some factors to consider
  • Cost of reclaiming cells
  • Cost of allocating cells
  • Storage overhead
  • How does the algorithm scale with residency?
  • Will user program be suspended during garbage
    collection?
  • Does an upper bound exist on the pause time?
  • Is locality of data structures maintained (or
    maybe even improved?)

9
Classes of Garbage Collection Algorithms
  • Direct Garbage Collectors a record is associated
    with each node in the heap. The record for node
    N indicates how many other nodes or roots point
    to N.
  • Indirect/Tracing Garbage Collectors usually
    invoked when a users request for memory fails
    because the free list is exhausted. The garbage
    collector visits all live nodes, and returns all
    other memory to the free list. If sufficient
    memory has been recovered from this process, the
    users request for memory is satisfied.

10
Quick Review Reference Counting
  • Every cell has an additional field the reference
    count. This field represents the number of
    pointers to that cell from roots or heap cells.
  • Initially, all cells in the heap are placed in a
    pool of free cells, the free list.

11
Reference Counting(Continued)
  • When a cell is allocated from the free list, its
    reference count is set to one.
  • When a pointer is set to reference a cell, the
    cells reference count is incremented by 1 if a
    pointer is to the cell is deleted, its reference
    count is decremented by 1.
  • When a cells reference count reaches 0, its
    pointers to its children are deleted and it is
    returned to the free list.

12
Reference Counting Example
1
0
1
0
2
0
1
0
1
13
Reference Counting Example (Continued)
1
2
1
0
1
1
14
Reference Counting Example (Continued)
1
2
1
0
1
1
15
Reference Counting Example (Continued)
1
2
1
0
1
1
0
1
0
16
Reference Counting Advantages and Disadvantages
  • Advantages
  • Garbage collection overhead is distributed.
  • Locality of reference is no worse than mutator.
  • Free memory is returned to free list quickly.

17
Reference Counting Advantages and
Disadvantages(Continued)
  • Disadvantages
  • High time cost (every time a pointer is changed,
    reference counts must be updated).
  • Storage overhead for reference counter can be
    high.
  • Unable to reclaim cyclic data structures.
  • If the reference counter overflows, the object
    becomes permanent.

18
Reference Counting Cyclic Data Structure -
Before
1
0
2
0
2
0
1
19
Reference Counting Cyclic Data Structure After
1
0
1
0
2
0
1
20
Deferred Reference Counting
  • Optimisation
  • Cost can be improved by special treatment of
    local variables.
  • Only update reference counters of objects on the
    stack at fixed intervals.
  • Reference counts are still affected from pointers
    from one heap object to another.

21
Quick Review Mark-Sweep
  • The first tracing garbage collection algorithm
  • Garbage cells are allowed to build up until heap
    space is exhausted (i.e. a user program requests
    a memory allocation, but there is insufficient
    free space on the heap to satisfy the request.)
  • At this point, the mark-sweep algorithm is
    invoked, and garbage cells are returned to the
    free list.

22
Mark-Sweep(Continued)
  • Performed in two phases
  • Mark phase identifies all live cells by setting
    a mark bit. Live cells are cells reachable from
    a root.
  • Sweep phase returns garbage cells to the free
    list.

23
Mark-Sweep Example
24
Mark-Sweep Advantages and Disadvantages
  • Advantages
  • Cyclic data structures can be recovered.
  • Tends to be faster than reference counting.

25
Mark-Sweep Advantages and Disadvantages(Continu
ed)
  • Disadvantages
  • Computation must be halted while garbage
    collection is being performed
  • Every live cell must be visited in the mark
    phase, and every cell in the heap must be visited
    in the sweep phase.
  • Garbage collection becomes more frequent as
    residency of a program increases.
  • May fragment memory.

26
Mark-Sweep Advantages and Disadvantages(Continu
ed)
  • Disadvantages
  • Has negative implications for locality of
    reference. Old objects get surrounded by new ones
    (not suited for virtual memory applications).
  • However, if objects tend to survive in clusters
    in memory, as they apparently often do, this can
    greatly reduce the cost of the sweep phase.

27
Mark-Compact Collection
  • Remedy the fragmentation and allocation problems
    of mark-sweep collectors.
  • Two phases
  • Mark phase identical to mark sweep.
  • Compaction phase marked objects are compacted,
    moving most of the live objects until all the
    live objects are contiguous.

28
Mark-Compact Advantages and Disadvantages(Conti
nued)
  • Advantages
  • The contiguous free area eliminates fragmentation
    problem. Allocating objects of various sizes is
    simple.
  • The garbage space is "squeezed out", without
    disturbing the original ordering of objects. This
    ameliorate locality.

29
Mark-Compact Advantages and Disadvantages(Conti
nued)
  • Disadvantages
  • Requires several passes over the data are
    required. "Sliding compactors" takes two, three
    or more passes over the live objects.
  • One pass computes the new location
  • Subsequent passes update the pointers to refer to
    new locations, and actually move the objects

30
Copying Garbage Collection
  • Like mark-compact, copying garbage collection
    does not really "collect" garbage.
  • Rather it moves all the live objects into one
    area and the rest of the heap is know to be
    available.
  • Copying collectors integrate the traversal and
    the copying process, so that objects need only be
    traversed once.
  • The work needed is proportional to the amount of
    live date (all of which must be copied).

31
Semispace Collector Using the Cheney Algorithm
  • The heap is subdivided into two contiguous
    subspaces (FromSpace and ToSpace).
  • During normal program execution, only one of
    these semispaces is in use.
  • When the garbage collector is called, all the
    live data are copied from the current semispace
    (FromSpace) to the other semispace (ToSpace).

32
Semispace Collector Using the Cheney Algorithm
A
B
C
D
FromSpace
ToSpace
33
Semispace Collector Using the Cheney Algorithm
C
D
A
B
A
B
C
D
FromSpace
ToSpace
34
Semispace Collector Using the Cheney
Algorithm(Continued)
  • Once the copying is completed, the ToSpace is
    made the "current" semispace.
  • A simple form of copying traversal is the Cheney
    algorithm.
  • The immediately reachable objects from the
    initial queue of objects for a breadth-first
    traversal.
  • A scan pointer is advanced through the first
    object location by location.
  • Each time a pointer into FromSpace is
    encountered, the referred-to-object is
    transported to the end of the queue and the
    pointer to the object is updated.

35
Cheney Algorithm Example
B
A
Root Nodes
B
F
A
B
C
A
E
D
C
A
B
C
D
C
D
A
B
E
A
B
C
D
E
F
36
Semispace Collector Using the Cheney Algorithm
(Continued)
  • Multiple paths must not be copied to tospace
    multiple times.
  • When an object is transported to tospace, a
    forwarding pointer is installed in the old
    version of the object.
  • The forwarding pointer signifies that the old
    object is obsolete and indicates where to find
    the new copy.

37
Copying Garbage Collection Advantages and
Disadvantages
  • Advantages
  • Allocation is extremely cheap.
  • Excellent asymptotic complexity.
  • Fragmentation is eliminated.
  • Only one pass through the data is required.

38
Copying Garbage Collection Advantages and
Disadvantages(Continued)
  • Disadvantages
  • The use of two semi-spaces doubles memory
    requirement needs
  • Poor locality. Using virtual memory will cause
    excessive paging.

39
Problems with Simple Tracing Collectors
  • Difficult to achieve high efficiency in a simple
    garbage collector, because large amounts of
    memory are expensive.
  • If virtual memory is used, the poor locality of
    the allocation/reclamation cycle will cause
    excessive paging.
  • Even as main memory becomes steadily cheaper,
    locality within cache memory becomes increasingly
    important.

40
Problems with Simple Tracing Collectors(Continued
)
  • With a simple semispace copy collector, locality
    is likely to be worse than mark-sweep.
  • The memory issue is not unique to copying
    collectors.
  • Any efficient garbage collection involves a
    trade-off between space and time.
  • The problem of locality is an indirect result of
    the use of garbage collection.

41
Incremental Tracing Collectors Overview
  • Introduction to Incremental Collectors
  • Coherence and Conservatism
  • Tricolor Marking
  • Write Barrier Algorithms
  • Bakers Read Barrier Algorithm

42
Incremental Tracing Collectors
  • Program (Mutator) and Garbage Collector run
    concurrently.
  • Can think of system as similar to two threads.
    One performs collection, and the other represents
    the regular program in execution.
  • Can be used in systems with real-time
    requirements. For example, process control
    systems.

43
Coherence Conservatism
  • Coherence A proper state must be maintained
    between the mutator and the collector.
  • Conservatism How aggressive the garbage
    collector is at finding objects to be
    deallocated.

44
Tricoloring
  • White Not yet traversed. A candidate for
    collection.
  • Black Already traversed and found to be live.
    Will not be reclaimed.
  • Grey In traversal process. Defining
    characteristic is that its children have not
    necessarily been explored.

45
The Tricolor Abstraction
46
Tricoloring Invariant
  • There must not be a pointer from a black object
    to a white object.

47
Violation of Coloring Invariant
A
A
B
C
B
C
D
D
Before
After
48
Steps in Violation
  • Read a pointer to a white object
  • Assign that pointer to a black object
  • Original pointer must be destroyed without
    collection system noticing.

49
Read Barrier
  • Barriers are essentially memory access detection
    systems.
  • We detect when any pointers to any white objects
    are read.
  • If a read to the pointer occurs, we conceptually
    color that object grey.

50
Write Barrier
  • When a pointer is written to an object, we record
    the write somehow.
  • The recorded write is dealt with at a later
    point.
  • Read vs. Write efficiency considerations.

51
Write Barrier Algorithms
  • Snapshot-at-beginning
  • Incremental update

52
Snapshot-at-beginning
  • Conceptually makes a copy-on-write duplication of
    the pointer graph.
  • Can be implemented with a simple write barrier
    that records pointer writes and adds the old
    addresses to a stack to be traversed later.

53
Snapshot-at-beginning Example
A
A
Pointer to D is now On stack
B
C
B
C
D
D
Stack
Before
After
54
Comments on Snapshot-at-beginning
  • Very conservative.
  • All overwritten pointer values are saved and
    traversed.
  • No objects can be freed while collection process
    is occurring.

55
Incremental Update Write-Barrier Algorithm
  • No copy of tree is made.
  • Catches overwrites of pointers that have been
    copied.
  • If a pointer is not copied before being written,
    it will be freed.
  • The object with the overwritten pointer is
    colored grey and the algorithm must search that
    node again at the end.

56
Incremental Update Example
A
A
B
C
B
C
D
D
Before
After
57
Comments on Incremental Update
  • Things that are freed during collection are far
    more likely to be collected than with the
    snapshot algorithm. (Less conservative)
  • Although the collector restarts the traversal in
    some places, it is guaranteed to do a full search
    and will eventually terminate.

58
Bakers Read Barrier Algorithms
  • Incremental Copying
  • Non-copying Algorithm (The Treadmill)

59
Incremental Copying
  • Variation of Copying Collector
  • Garbage collection cycle begins with an atomic
    flip.
  • All objects directly pointed to by the root are
    copied into tospace.

60
Read Barrier in Incremental Copying
  • Whenever an object is read that is not already in
    ToSpace, the read barrier catches that and copies
    the object over to ToSpace at that point.
  • Normal background scavenging occurs
    simultaneously to ensure that all objects are
    traversed and reclamation can occur.

61
Incremental Copying Example
A
B
D
A
B
C
C
D
E
FromSpace
ToSpace
Atomic Flip, then a read to D occurs
62
Comments on Read Barrier
  • If implemented in software can be quite slow due
    to numerous reads to heap.
  • Specialized hardware is available on some unique
    machines that allow this type of tracing to be
    done quickly.

63
Bakers Incremental Non-Copying Algorithm
  • Doubly Linked Lists
  • New area for allocations since started collection
  • To/From spaces
  • Free list

64
Example - Allocation
  • Take an object from the free list and move it to
    the new list.

65
Example - Scanning
  • Searching nodes in ToSpace for references to
    objects in FromSpace.
  • When found, object is unlinked in FromSpace and
    is linked in ToSpace.

66
Treadmill Workings
  • When starting collection cycle
  • New list is empty
  • From list contains all New and To objects from
    last cycle.
  • Collection proceeds and scanning and allocation
    are performed.
  • When finished
  • From list is merged with Free list.

67
Comments on Treadmill
  • As in Incremental Copying, the garbage found in
    the FromSpace is reclaimed in constant time.
  • Conservative with new objects
  • Conservative also in that reached objects will
    not be removed even if they become garbage before
    scan ends.

68
Incremental Collectors Summary
  • Incremental Tracing Collectors
  • Tricolor Marking and Invariant
  • Read and Write Barriers
  • Snapshot-at-beginning
  • Incremental Update
  • Bakers Incremental Copying
  • Bakers Non-copying (Treadmill)

69
Generational Garbage Collection
  • Attempts to address weaknesses of simple tracing
    collectors such as mark-sweep and copying
    collectors
  • All active data must be marked or copied.
  • For copying collectors, each page of the heap is
    touched every two collection cycles, even though
    the user program is only using half the heap,
    leading to poor cache behavior and page faults.
  • Long-lived objects are handled inefficiently.

70
Generational Garbage Collection(Continued)
  • Generational garbage collection is based on the
    generational hypothesis
  • Most objects die young.
  • As such, concentrate garbage collection efforts
    on objects likely to be garbage young objects.

71
Generational Garbage Collection Object Lifetimes
  • When we discuss object lifetimes, the amount of
    heap allocation that occurs between the objects
    birth and death is used rather than the wall
    time.
  • For example, an object created when 1Kb of heap
    was allocated and was no longer referenced when 4
    Kb of heap data was allocated would have lived
    for 3Kb.

72
Generational Garbage Collection Object
Lifetimes(Continued)
  • Typically, between 80 and 98 percent of all
    newly-allocated heap objects die before another
    megabyte has been allocated.

73
Generational Garbage Collection(Continued)
  • Objects are segregated into different areas of
    memory based on their age.
  • Areas containing newer objects are garbage
    collected more frequently.
  • After an object has survived a given number of
    collections, it is promoted to a less frequently
    collected area.

74
Generational Garbage Collection Example
C
B
A
Root Set
S
Memory Usage
Memory Usage
Old Generation
New Generation
75
Generational Garbage Collection
Example(Continued)
R
C
B
A
Root Set
S
Memory Usage
Memory Usage
Old Generation
New Generation
76
Generational Garbage Collection
Example(Continued)
R
D
C
B
A
Root Set
S
Memory Usage
Memory Usage
Old Generation
New Generation
77
Generational Garbage Collection
Example(Continued)
  • This example demonstrates several interesting
    characteristics of generational garbage
    collection
  • The young generation can be collected
    independently of the older generations (resulting
    in shorter pause times).
  • An intergenerational pointer was created from R
    to D. These pointers must be treated as part of
    the root set of the New Generation.
  • Garbage collection in the new generation result
    in S becoming unreachable, and thus garbage.
    Garbage in older generations (sometimes called
    tenured garbage) can not be reclaimed via garbage
    collections in younger generations.

78
Generational Garbage Collection Implementation
  • Usually implemented as a copying collector, where
    each generation has its own semispace

FromSpace
FromSpace
ToSpace
ToSpace
Old Generation
New Generation
79
Generational Garbage Collection Issues
  • Choosing an appropriate number of generations
  • If we benefit from dividing the heap into two
    generations, can we further benefit by using more
    than two generations?
  • Choosing a promotion policy
  • How many garbage collections should an object
    survive before being moved to an older generation?

80
Generational Garbage Collection
Issues(Continued)
  • Tracking intergenerational pointers
  • Inter-generational pointers need to be tracked,
    since they form part of the root set for younger
    generations.
  • Collection Scheduling
  • Can we attempt to schedule garbage collection in
    such a way that we minimize disruptive pauses?

81
Generational Garbage Collection Multiple
Generations
Generation 1
Generation 2
Generation 3
Generation 4
82
Generational Garbage Collection Multiple
Generations(Continued)
  • Advantages
  • Keeps youngest generations size small.
  • Helps address mistakes made by the promotion
    policy by creating more intermediate generations
    that still get garbage collected fairly
    frequently.
  • Disadvantages
  • Collections for intermediate generations may be
    disruptive.
  • Tends to increase number of inter-generational
    pointers, increasing the size of the root set for
    younger generations.
  • Most generational collectors are limited to just
    two or three generations.

83
Generational Garbage Collection Promotion
Policies
  • A promotion policy determines how many garbage
    collections cycles (the cycle count) an object
    must survive before being advanced to the next
    generation.
  • If the cycle count is too low, objects may be
    advanced too fast if too high, the benefits of
    generational garbage collection are not realized.

84
Generational Garbage Collection Promotion
Policies(Continued)
  • With a cycle count of just one, objects created
    just before the garbage collection will be
    advanced, even though the generational hypothesis
    states they are likely to die soon.
  • Increasing the cycle count to two denies
    advancement to recently created objects.
  • Under most conditions, it increasing the cycle
    count beyond two does not significantly reduce
    the amount of data advanced.

85
Generational Garbage Collection
Inter-generational Pointers
  • Inter-generational pointers can be created in two
    ways
  • When an object containing pointers is promoted to
    an older generation.
  • When a pointer to an object in a newer generation
    is stored in an object.
  • The garbage collector can easily detect
    promotion-caused inter-generational pointers, but
    handling pointer stores is a more complicated
    task.

86
Generational Garbage Collection
Inter-generational Pointers
  • Pointer stores can be tracked via the use of a
    write barrier
  • Pointer stores must be accompanied by extra
    bookkeeping instructions that let the garbage
    collector know of pointers that have been
    updated.
  • Often implemented at the compiler level.

87
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • However, write barriers only provide a
    conservative estimation of live intergenerational
    pointers

Root Set
Old Generation
New Generation
88
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • Tracking inter-generational pointers are often
    the largest cost of generational garbage
    collection.
  • 1 percent of a typical Lisp programs total
    instruction count are pointer stores. If a write
    barrier adds 10 instructions to a pointer store,
    overall performance will drop by 10 percent.

89
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • Entry Tables
  • Pointers from older generations point indirectly
    to younger generations via an entry table

Generation 2
Generation 1
Generation 3
Entry Table
Entry Table
90
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • Entry Table Advantages
  • When a younger generation is collected, only the
    entry table for that generation needs to be
    scanned.
  • Entry Table Disadvantages
  • Entry table may contain several entries to the
    same object, making scans of the object table
    proportional to the number of pointer stores
    rather than to the number of inter-generational
    pointers.
  • High overhead because of extra level of
    indirection.

91
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • Remembered Sets
  • The write barrier checks to see if a pointer
    being stored in an old objects points to an
    object in a newer generation. If so, the address
    of the old object is added to the remembered set
    (if that object is not already in the set).

92
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • Remembered Sets (Continued)

New Generation
Old Generation
Remembered Set
93
Generational Garbage Collection
Inter-generational Pointers(Continued)
  • Remembered Sets Advantages
  • Scanning is proportional to the number of
    stored-into objects, not the number of store
    operations.
  • Remembered Sets Disadvantages
  • Pointer store checking can be expensive.

94
Generational Garbage Collection Collection
Scheduling
  • Generational garbage collection aims to reduce
    pause times. When should these (hopefully short)
    pause times occur?
  • Two strategies exist
  • Hide collections when the user is least likely to
    notice a pause, or
  • Trigger efficient collections when there is
    likely to be lots of garbage to collect.

95
Generational Garbage Collection Advantages
  • In practice it has proven to be an effective
    garbage collection technique.
  • Minor garbage collections are performed quickly.
  • Good cache and virtual memory behavior.

96
Generational Garbage Collection Disadvantages
  • Performs poorly if any of the main assumptions
    are false
  • That objects tend die young.
  • That there are relatively few pointers from old
    objects to young ones.
  • Frequent pointer writes to older generations will
    increase the cost of the write barrier, and
    possibly increase the size of the root set for
    younger generations.

97
Garbage Collection Summary
Method Conservatism Space Time Fragmentation Locality
Mark Sweep Major Basic 1 traversal heap scan Yes Fair
Mark Compact Major Basic Many passes of heap No Good
Copying Major Two Semispaces 1 traversal No Poor
Reference Counting No Reference count field Constant per Assignment Yes Very Good
Deferred Reference Counting Only for stack variables Reference Count Field Constant per Assignment Yes Very Good
Incremental Varies depending on algorithm Varies Can be Guaranteed Real-Time Varies Varies
Generational Variable Segregated Areas Varies with number of live objects in new generation Yes (Non-Copying) No (Copying) Good
Tracing
Incremental
98
Garbage Collection Conclusions
  • Relieves the burden of explicit memory allocation
    and deallocation.
  • Software module coupling related to memory
    management issues is eliminated.
  • An extremely dangerous class of bugs is
    eliminated.

99
Garbage Collection Conclusions(Continued)
  • Zorns study in 1989/93 compared garbage
    collection to explicit deallocation
  • Non-generational
  • Between 0 and 36 more CPU time.
  • Between 40 and 280 more memory.
  • Generational garbage collection
  • Between 5 to 20 more CPU time.
  • Between 30 and 150 more memory.
  • Wilson feels these numbers can be improved, and
    they are also out of date.
  • A well implemented garbage collector will slow a
    program down by approximately 10 percent relative
    to explicit heap deallocation.

100
Garbage Collection Conclusions(Continued)
  • Despite this cost, garbage collection a feature
    in many widely used languages
  • Lisp (1959)
  • Perl (1987)
  • Java (1995)
  • C (2001)
  • Microsofts Common Language Runtime (2002)

101
Garbage Collection Pointers
  • Heap of fish applet (Mark and Sweep garbage
    collection example)
  • http//www.artima.com/insidejvm/applets/HeapOfFish
    .html
  • Java HotSpot Garbage Collection Strategies
  • http//developer.java.sun.com/developer/technicalA
    rticles/Networking/HotSpot/
  • The Memory Management Reference
  • http//www.memorymanagement.org/
  • Uniprocessor Garbage Collection Techniques
    (Wilson)
  • http//www.cs.ualberta.ca/duane/courses/425-52
    5/WilsonACMDraft.pdf
  • Garbage Collection Algorithms for Automatic
    Dynamic Memory Management
  • (Richard Jones and Rafael Lins)

102
Questions?
  • If you have any questions, please feel free to
  • e-mail one of us
  • Patrick Earl patrick_at_cs.ualberta.ca
  • Simon Leonard sleonard_at_cs.ualberta.ca
  • Jack Newton newton_at_cs.ualberta.ca
Write a Comment
User Comments (0)
About PowerShow.com