When to use a garbage collector?

[Carlo Milanesi]

Hello,
traditionally, in C++, dynamically allocated memory has been
managed explicitly by calling "delete" in the application code.

Now, in addition to the standard library strings, containers, and
auto_ptrs, gurus suggest that may be better to use a reference-counted
smart pointer, or a garbage-collector.

But in which cases it is better to use one technique and in which cases
another? IOW, which is the design criterion?

And if, after having completed a working system, a technique would
result more performing than another, or better for other reasons, is it
advisable to change the memory management strategy of that working system?

 

[Krice]

But in which cases it is better to use one technique and in which cases
another?

It's probably a question that has no answer. But I like the
old school delete, because it's transparent, you know exactly
what it's doing and when.

 

[Jerry Coffin]

[ ... ]

Now, in addition to the standard library strings, containers, and
auto_ptrs, gurus suggest that may be better to use a reference-counted
smart pointer, or a garbage-collector.

_Some_ gurus -- there are others (who I think qualify as gurus) who are
opposed to the use of garbage collection to varying degrees, or at least
are of the view that its use should be restricted to specific
situations.

But in which cases it is better to use one technique and in which cases
another? IOW, which is the design criterion?

The cost of garbage collection varies based on the type of collector
used. Nearly every garbage collector has a "mark" phase, in which
"live" objects (i.e. those that are still accessible) are marked as
being in use. The cost of this phase is normally about linear on the
number of objects currently accessible to the program.

After that, different garbage collectors work in a number of different
ways. A copying collector copies all those live objects into a
contiguous space in memory, leaving another contiguous free space. This
makes allocations extremely cheap, and the cost of the collection as a
whole is also linear on the number of objects that are accessible.

Other collectors leave the "live" objects where they are, and create
free blocks of all the contiguous chunks of memory in the heap not
currently occupied by live objects. In this case, the cost of the
collection part tends to relate most closely to the number of contiguous
chunks of free memory in the heap.

On the other side, life isn't simple either. Manual memory management
tends to have costs associated most closely with the number of
allocations and frees used. This cost can be mitigated (drastically) in
certain cases, such as allocating a large number of objects of identical
size, or releasing a large number of objects all at the same, if the
allocator is written to allow them to be released together rather than
individually.

That only begins to tell the real story though: much of the cost of
manual allocation/deletion arises when objects are copied. A garbage
collector can (and does) keep track of different pointers that refer to
an object, and only deletes the object when all pointers that give
access to the object are gone. This makes it easy to keep a single
object, and create new pointers/references to that object whenever
needed.

With manual memory management, it's far more common to duplicate the
entire object, so each time the object is used, there's a separate
instance of the object to look at, and each instance has exactly one
owner that's responsible for deleting the object when it's no longer
needed. Allocating space to hold extra copies of the object, and copying
the relevant data into each copy, can take a considerable amount of
time. With a GC in place, you can usually avoid this copying by just
passing around pointers and everything shares access to that one object.
OTOH, when/if you need to copy the object anyway (e.g. if one copy will
be modified to become different from the other), this does little good.

As such, a tremendous amount depends upon things like: 1) what you're
allocating dynamically, 2) how you're using the dynamically allocated
objects, and 3) the degree to which objects you'd copy with manual
memory management can be shared in the presence of GC. All of these (and
more) depend on the application and design, not just the memory
management itself.

 

[Stefan Ram]

Allocating space to hold extra copies of the object, and copying
the relevant data into each copy, can take a considerable amount of
time. With a GC in place, you can usually avoid this copying by just
passing around pointers and everything shares access to that one object.

»There were two versions of it, one in Lisp and one in
C++. The display subsystem of the Lisp version was faster.
There were various reasons, but an important one was GC:
the C++ code copied a lot of buffers because they got
passed around in fairly complex ways, so it could be quite
difficult to know when one could be deallocated. To avoid
that problem, the C++ programmers just copied. The Lisp
was GCed, so the Lisp programmers never had to worry about
it; they just passed the buffers around, which reduced
both memory use and CPU cycles spent copying.«

<[email protected]>

»A lot of us thought in the 1990s that the big battle would
be between procedural and object oriented programming, and
we thought that object oriented programming would provide
a big boost in programmer productivity. I thought that,
too. Some people still think that. It turns out we were
wrong. Object oriented programming is handy dandy, but
it's not really the productivity booster that was
promised. The real significant productivity advance we've
had in programming has been from languages which manage
memory for you automatically.«

http://www.joelonsoftware.com/articles/APIWar.html

»[A]llocation in modern JVMs is far faster than the best
performing malloc implementations. The common code path
for new Object() in HotSpot 1.4.2 and later is
approximately 10 machine instructions (data provided by
Sun; see Resources), whereas the best performing malloc
implementations in C require on average between 60 and 100
instructions per call (Detlefs, et. al.; see Resources).
And allocation performance is not a trivial component of
overall performance -- benchmarks show that many
real-world C and C++ programs, such as Perl and
Ghostscript, spend 20 to 30 percent of their total
execution time in malloc and free -- far more than the
allocation and garbage collection overhead of a healthy
Java application (Zorn; see Resources).«

http://www-128.ibm.com/developerworks/java/library/j-jtp09275.html?ca=dgr-jw22JavaUrbanLegends

»Perhaps the most important realisation I had while developing
this critique is that high level languages are more important
to programming than object-orientation. That is, languages
which have the attribute that they remove the burden of
bookkeeping from the programmer to enhance maintainability and
flexibility are more significant than languages which just
add object-oriented features. While C++ adds object-orientation
to C, it fails in the more important attribute of being high
level. This greatly diminishes any benefits of the
object-oriented paradigm.«

http://burks.brighton.ac.uk/burks/pcinfo/progdocs/cppcrit/index005.htm

 

[AnonMail2005]

Hello,
     traditionally, in C++, dynamically allocated memory has been
managed explicitly by calling "delete" in the application code.

Now, in addition to the standard library strings, containers, and
auto_ptrs, gurus suggest that may be better to use a reference-counted
smart pointer, or a garbage-collector.

But in which cases it is better to use one technique and in which cases
another? IOW, which is the design criterion?

And if, after having completed a working system, a technique would
result more performing than another, or better for other reasons, is it
advisable to change the memory management strategy of that working system?

I almost always use smart pointers for objects. In any moderately
sized
system, the complexity of figuring out when to delete something is
just
not worth it. With smart pointers, you never have to delete.

Of course, memory managed *within* a class can be handled with raw
pointers.

So I don't see a real choie between using deletes and using smart
pointers
unless there is a special circumstance or platform.

We have yet to move or consider other forms of garbage collectors such
as that used in Java. Call me old fashioned .

 

[acehreli]

It's probably a question that has no answer. But I like the
old school delete, because it's transparent, you know exactly
what it's doing and when.

I don't think so. Can you tell whether the delete here will be called:

void foo()
{
A * a = new A(/* ... */);
/* code that may throw today or some time in the future possibly
after some code change */
delete a;
}

To the OP: never deallocate resources explicitly in code. That would
not be "traditional," rather "developmental." I used to do that
when I didn't know better. :/

Ali

 

[acehreli]

On Jun 10, 11:30 am, Carlo Milanesi <[email protected]>
wrote:

I almost always use smart pointers for objects. In any moderately
sized
system, the complexity of figuring out when to delete something is
just
not worth it.

The problem is, that complex code hoping to figure out when to
'delete' an object may never be executed. The non-throwing lines of
code of today can suddenly start possibly throwing in the future by
code changes, and the explicit delete statement may never be executed.

Of course, memory managed *within* a class can be handled with raw
pointers.

Unfortunately that statement must be qualified further: you are
talking about classes that manage a single object, right? Because the
following class fits your description, but is not exception-safe and
cannot manage the memory it hopes to manage:

class HopefulManager
{
One * one_;
Two * two_;

public:

HopefulManager()
:
one_(new One()),
two_(new Two()) // <- 1, <- 2
{
// some other code that may throw <- 3
}

~HopefulManager()
{
delete two_;
delete one_;
}
};

The three places that may cause resource leaks are:

1) new may throw, one_'s object is leaked

2) Two() may throw, one_'s object is leaked

3) Any line is the constructor may throw, one_'s and two_'s objects
are leaked

So I don't see a real choie between using deletes and using smart
pointers
unless there is a special circumstance or platform.

What I said above is regardless of special circumstances or platforms.
Pure C++...

Ali

 

[Jerry Coffin]

»There were two versions of it, one in Lisp and one in
C++. The display subsystem of the Lisp version was faster.
There were various reasons, but an important one was GC:
the C++ code copied a lot of buffers because they got
passed around in fairly complex ways, so it could be quite
difficult to know when one could be deallocated. To avoid
that problem, the C++ programmers just copied. The Lisp
was GCed, so the Lisp programmers never had to worry about
it; they just passed the buffers around, which reduced
both memory use and CPU cycles spent copying.«

<[email protected]>

Intentionally or otherwise, I suspect your post is likely to generate
fare more heat than light. Most of it is unsupported assertions, and
none of it is from anybody who appears to deserve the title of "guru",
at least with respect to C++ (and IMO, probably not in any other respect
either).

The first quote appears to be purely apocryphal -- an unsupported
statement from somebody posting under a pseudonym, about software of
unknown origin written by people of unknown skills.

Joel Spolsky spends a lot of time writing about software, but his
credentials seem questionable at best. In particular, I've seen nothing
to give a really strong indication that he's much of a programmer
(himself) at all.

IBM, of course, has a great deal of collective knowledge about
programming -- but the bit you quote is written with the specific intent
of promoting Java. It's been discussed here before, and at very best
it's misleading when applied to more than the very specific domain about
which it's written.

Finally we get yet another reference to Ian Joyner's "Critique of C++."
IMO, there should be something similar to Godwin's law relating to
anybody who quotes (any part of) this. First of all, it has nothing to
do with the C++ of today, or anytime in the last decade or more. As of
the first edition, some (a small fraction) was reasonably accurate about
the C++ of the time -- but the updates in his second and third editions
were insufficient to keep the relevant to the C++ of their times, and
the third edition still predates the original C++ standard by a couple
of years. With respect to the C++ of today, it varies from irrelevant to
misleading to downright false. Second, a great deal of it was misleading
when it was originally written. Third, nearly all the rest of it was
downright false when written.

When you get down to it, despite being umpteen pages long, the
criticisms in this paper that have at least some degree of validity with
respect to current C++ can be summarized as:

1) member functions should be virtual by default.
2) C++ should have Concepts [JVC: C++ 0x will].
3) Unified syntax for "." and "->" would be nice.
4) "static" is overloaded in too many (confusing) ways.
5) Modules would be better than headers.
6) Support for DbC would have some good points.

When you get down to it, however, it would be much easier to summarize
his critique in a single sentence: "C++ isn't Eiffel." Many of his
individual arguments aren't really supported at all -- they're simply
statements that C++ must be wrong because it's different from Eiffel.

Don't get me wrong: my previous statement that GC is favored for some
situations under some circumstances still stands -- but IMO, none of
these quotes provides any real enlightenment. Quite the contrary, the
quote from IBM means _almost_ nothing, and the other three (between
them) mean far less still.

 

[Fran]

The problem is, that complex code hoping to figure out when to
'delete' an object may never be executed. The non-throwing lines of
code of today can suddenly start possibly throwing in the future by
code changes, and the explicit delete statement may never be executed.

Is there any chance that C++0x will give us mandatory checking of
throw() clauses in function definitions? That would enable the
compiler to warn about leaks when exceptions might happen between
calls to new and delete.

 

[James Kanze]

Intentionally or otherwise, I suspect your post is likely to
generate fare more heat than light. Most of it is unsupported
assertions, and none of it is from anybody who appears to
deserve the title of "guru", at least with respect to C++ (and
IMO, probably not in any other respect either).

You noticed that too.

The first quote appears to be purely apocryphal -- an
unsupported statement from somebody posting under a pseudonym,
about software of unknown origin written by people of unknown
skills.

The first quote is probably the one which does correspond most
to practical reality; I seem to recall a similar statement being
made by Walter Bright (who certainly does qualify as a C++
guru). But of course, it doesn't have to be that way.

Joel Spolsky spends a lot of time writing about software, but
his credentials seem questionable at best. In particular, I've
seen nothing to give a really strong indication that he's much
of a programmer (himself) at all.

Another case of "those who can, do; those who can't teach (or
write articles)".

IBM, of course, has a great deal of collective knowledge about
programming -- but the bit you quote is written with the
specific intent of promoting Java. It's been discussed here
before, and at very best it's misleading when applied to more
than the very specific domain about which it's written.

Finally we get yet another reference to Ian Joyner's "Critique
of C++." IMO, there should be something similar to Godwin's
law relating to anybody who quotes (any part of) this. First
of all, it has nothing to do with the C++ of today, or anytime
in the last decade or more. As of the first edition, some (a
small fraction) was reasonably accurate about the C++ of the
time -- but the updates in his second and third editions were
insufficient to keep the relevant to the C++ of their times,
and the third edition still predates the original C++ standard
by a couple of years. With respect to the C++ of today, it
varies from irrelevant to misleading to downright false.
Second, a great deal of it was misleading when it was
originally written. Third, nearly all the rest of it was
downright false when written.

When you get down to it, despite being umpteen pages long, the
criticisms in this paper that have at least some degree of
validity with respect to current C++ can be summarized as:

1) member functions should be virtual by default.

Which is just wrong, at least from a software engineering point
of view.

2) C++ should have Concepts [JVC: C++ 0x will].
3) Unified syntax for "." and "->" would be nice.

I don't think I agree with this one, either.

4) "static" is overloaded in too many (confusing) ways.

The price we pay for C compatibility.

5) Modules would be better than headers.

I don't think anyone could disagree with that one. Of course,
just about everyone has a different definition of what they mean
by "modules".

6) Support for DbC would have some good points.

Interestingly, I think that C++ today has the best support of
any language, although it's not automatic, and many programmers
fail to use it.

When you get down to it, however, it would be much easier to
summarize his critique in a single sentence: "C++ isn't
Eiffel." Many of his individual arguments aren't really
supported at all -- they're simply statements that C++ must be
wrong because it's different from Eiffel.

Don't get me wrong: my previous statement that GC is favored
for some situations under some circumstances still stands --
but IMO, none of these quotes provides any real enlightenment.
Quite the contrary, the quote from IBM means _almost_ nothing,
and the other three (between them) mean far less still.

I don't think that there is complete consensus among the gurus
as to when garbage collection would be appropriate. I would be
very suspicious, however, of anyone who claimed that it is
always appropriate, or never appropriate. That it's not
available in the standard toolkit is a definite flaw in the
language, but requiring it to be used in every case would
probably be even worse (but I don't think anyone has ever
proposted that).

 

[Pascal J. Bourguignon]

Hello,
traditionally, in C++, dynamically allocated memory has been
managed explicitly by calling "delete" in the application code.

Now, in addition to the standard library strings, containers, and
auto_ptrs, gurus suggest that may be better to use a reference-counted
smart pointer, or a garbage-collector.

But in which cases it is better to use one technique and in which
cases another? IOW, which is the design criterion?

Reference counted smart pointers: never. They leak memory as soon as
you have bidirectionnal associations or cycles in your data
structures.

Garbage collectors: always. There are even real-time garbage
collectors, if you have real-time constraints.

And if, after having completed a working system, a technique would
result more performing than another, or better for other reasons, is
it advisable to change the memory management strategy of that working
system?

Well, usually garbage collectors give better performance.
http://www.jwz.org/doc/gc.html
But of course it depends on the application and datasets. It might be
easier to change the garbage collection strategy, selecting one more
adapted to the application, than to change the application to use
another memory management style. On the other hand, if you write your
code with reference counting, it is easy enough to disable reference
counting and fall back to garbage collection.

 

[dizzy]

Hello,
traditionally, in C++, dynamically allocated memory has been
managed explicitly by calling "delete" in the application code.

I think this is a misuderstanding. If by "traditionally" you mean C++ code
written until about 1998 maybe so, but the current C++ standard along with
auto_ptr<> and various third-party shared_ptr/counted_ptr imlementations
exist since at least 10 years now. Not to mention C++0x is just around the
corner and it will change the way we think C++, so we really need to drop
tradition and use best solutions as offered by a modern C++ implementation.

You probably didn't ment that by traditionally, in that case I'm stating the
above for those that do mean it that way

Now, in addition to the standard library strings, containers, and
auto_ptrs, gurus suggest that may be better to use a reference-counted
smart pointer, or a garbage-collector.

I'm no guru but I'll state my oppinion based on my experience.

There are many "problems" with using blind pointers. All IMO derive from the
fact that a pointer is just too semantically rich. Take for example a C
complex program, you will see pointers used as such:
- 90% of cases used as a reference (that is used to pass by reference
arguments and they actually cannot logically be NULL inside the function
receiving it but no explicit syntax states that)
- 9% of times used to signal optionality of value (that is, if the value
exists then it's != 0 or if it doesn't it is 0)
- 1% used for pointer arithmetic and other advanced pointer semantics

Ok, those numbers are obviously exagerated but you get the idea. Using a
pointer is like designing a class type that has all the program logic in it
(a monster with a huge number of memer functins and data with no invariants
to keep). Good design dictates that classes should be made as much
specialized as you can, to do one simple thing and do it well (and maintain
the invariants of that).

Same with pointers, you need to pass by references something around use a
reference (for which you know it can't be NULL since in order to initialize
a reference to something NULL you would have had to dereference NULL which
is UB anyway). You need to signal optional value, use something like
boost:ptional (many even very experienced C++ programmers still prefer
pointers for that). You need a pointer to own some memory and delete the
memory on end scope of the pointer, use an scoped_ptr (maybe an auto_ptr if
you prefer only std code or if you need the auto_ptr move semantics). And
so on.

All this specialization helps your program by moving into the C++ compile
time type system alot of your program logic (the fact that there is no
operator++ on auto_ptr<> means you can't do pointer arithmetic by mistake
on auto_ptr<> values, signaled by a compile time error) thus resulting into
a less error prone program not to mention easier to read by a reader that
knows what all those "alternative pointer types" do (since she won't have
to see how do you use a pointer, she sees an auto_ptr and knows already
some facts about your usage of it).

But in which cases it is better to use one technique and in which cases
another? IOW, which is the design criterion?

I think in general you should almost never use pointers. First go through
these alternative solutions and see which best fits your needs: C++
references, boost:ptional (optional is interesting also because you can
use it to build a sort of pointer value that can be either NULL or point
validly if for example you make an boost:ptional<T&>), auto_ptr /
scoped_ptr / shared_ptr / weak_ptr.

About gc I can't say much since I haven't used it in C++ (only in Java as it
was forced in). Since for my kind of development I deal with alot of
resources for which RAII and scoped objects map very well I have no need of
gc.

And if, after having completed a working system, a technique would
result more performing than another, or better for other reasons, is it
advisable to change the memory management strategy of that working system?

I think there are some situations in which a gc should perform better than
probably all those solutions above, maybe someone more experienced with
using gc's can provide an example (because all examples I come with right
now I also find them a solution C++ by doing a custom allocator).

 

[dizzy]

Reference counted smart pointers: never. They leak memory as soon as
you have bidirectionnal associations or cycles in your data
structures.

By that logic you mean he will always have bidirectional associations or
cycles in his data structures (thus NEVER use shared_ptr). In my years of
C++ I've had that very rare and when I did, I used weak_ptr to break the
cycle. How often do you have bidirectional associations in your data
structures? In thos projects that you have, which percent of the data
structures from the project has cycles?

Garbage collectors: always. There are even real-time garbage
collectors, if you have real-time constraints.

gc's are no silver bullet. They may be good in some scenarios but I don't
think they are good in any situation. Plus memory management is just a
small part of resource management in a C++ program (at least in my
programs).

Well, usually garbage collectors give better performance.
http://www.jwz.org/doc/gc.html
But of course it depends on the application and datasets.

Ah so then you contradict your previous "Garbage collectors: always".

 

[Pascal J. Bourguignon]

By that logic you mean he will always have bidirectional associations or
cycles in his data structures (thus NEVER use shared_ptr). In my years of
C++ I've had that very rare and when I did, I used weak_ptr to break the
cycle. How often do you have bidirectional associations in your data
structures? In thos projects that you have, which percent of the data
structures from the project has cycles?

Often enough. But the main point is of course, if your language
doesn't allow you to express some ideas easily, then you won't try to
express those ideas. If having circular references in C++ is a PITA,
then we will try very hard to avoid them. (And thus, burning a lot of
wetware cycles that would be better allocated to resolving the true
problems, instead of these technicalities).

Ah so then you contradict your previous "Garbage collectors: always".

Not really, in the following sentence you cut out, I explained that if
the currrent garbage collection algorithm wasn't good enough for your
application, it would be better to change this garbage collection
algorithm for another one more adapted to your particular
circumstances, rather than going back to manage memory manually.

 

[dizzy]

Is there any chance that C++0x will give us mandatory checking of
throw() clauses in function definitions? That would enable the
compiler to warn about leaks when exceptions might happen between
calls to new and delete.

Why is there such a need? Always assume anything may throw and code
accordingly. In the exceptional cases where writing exception safe code is
not possible (making the code expensive or error prone) I'm sure you can
find a solution (like std::stack has for top()/pop()).

 

[Krice]

void foo()
{
A * a = new A(/* ... */);
/* code that may throw today or some time in the future possibly
after some code change */
delete a;

}

Throw what? A ball?

 

[Jerry Coffin]

[ ... ]

The first quote is probably the one which does correspond most
to practical reality; I seem to recall a similar statement being
made by Walter Bright (who certainly does qualify as a C++
guru). But of course, it doesn't have to be that way.

Right -- my point wasn't that the quote was wrong, only that it didn't
really add much. If somebody disagreed with (essentially) the same point
when I said it probably wouldn't find much in this to convince them (of
anything).

Another case of "those who can, do; those who can't teach (or
write articles)".

....except that most of the people I can think of who write specifically
about C++ really _can_ write code, and most of them clearly _do_, and as
a rule do it quite well at that. The only prominent exception would be
Scott Meyers, who's pretty open about the fact that he consults about
C++, teaches C++, but does NOT really write much C++ at all. OTOH, I'm
pretty sure that if he really needed (or wanted to) he could write code
quite nicely as well -- though given his talents as a teacher, I think
it would be rather a waste if he spent his time that way.

Others, however (e.g. David Abrahams, Andrei Alexandrescu, Andrew
Koenig, Herb Sutter) who write about C++, also appear to write a fair
amount of code, and mostly do it quite well at that (and no, I'm not
claiming to be such a guru that I'm in a position to rate the experts,
or anything like that...)

[ ... ]

Which is just wrong, at least from a software engineering point
of view.

Right -- I don't mean to imply that all these are correct, or anything
like that -- I just mean that:

1) they're clearly enough defined to be fairly sure what he's saying.
2) they aren't obviously obsolete.
3) They aren't simply of the form: "Eiffel does it differently."

They're points that can be discussed intelligently, their strengths and
weaknesses can be examined, etc. They're not necessarily right, but at
least you can define (to at least some degree) what it means for them to
be right or wrong.

[ ... ]

I don't think that there is complete consensus among the gurus
as to when garbage collection would be appropriate. I would be
very suspicious, however, of anyone who claimed that it is
always appropriate, or never appropriate. That it's not
available in the standard toolkit is a definite flaw in the
language, but requiring it to be used in every case would
probably be even worse (but I don't think anyone has ever
proposted that).

I'm not sure it really needs to be part of the standard library, but I
do think it would be a good thing to tighten up the language
specification to the point that almost any known type of GC could be
included without leading to undefined behavior -- but we've been over
that before...

 

[Pascal J. Bourguignon]

I'm not sure it really needs to be part of the standard library, but I
do think it would be a good thing to tighten up the language
specification to the point that almost any known type of GC could be
included without leading to undefined behavior

Well said!

-- but we've been over
that before...

Ok.

 

[acehreli]

Throw what? A ball?

No, an exception; unless you are imagining a type named "ball" of
course.

I can't believe you are serious; you must have forgotten the smiley...
Seriously though, if you really didn't know what I meant with "throw,"
you should learn about exceptions.

Ali

 

[Krice]

Seriously though, if you really didn't know what I meant with "throw,"
you should learn about exceptions.

Exceptions are not logical. If construction of the object
fails then what? The program fails also, usually. I never
check anything, not since they invented exceptions, so
I'm assuming that there are no exceptions