...one of the most highly
regarded and expertly designed C++ library projects in the
world.
— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards
Boost.Intrusive containers offer speed improvements compared to non-intrusive containers primarily because:
This section will show performance tests comparing some operations on boost::intrusive::list
and
std::list
:
push_back
and container destruction will show the overhead associated with memory allocation/deallocation.
reverse
member function
will show the advantages of the compact memory representation that can be
achieved with intrusive containers.
sort
and write access
tests will show the advantage of intrusive containers minimizing memory accesses
compared to containers of pointers.
Given an object of type T
,
boost::intrusive::list<T>
can replace std::list<T>
to
avoid memory allocation overhead, or it can replace std::list<T*>
when
the user wants containers with polymorphic values or wants to share values
between several containers. Because of this versatility, the performance tests
will be executed for 6 different list types:
itest_class
,
each one with a different linking policy (normal_link
,
safe_link
, auto_unlink
). The itest_class
objects will be tightly packed in a std::vector<itest_class>
object.
std::list<test_class>
,
where test_class
is exactly
the same as itest_class
,
but without the intrusive hook.
std::list<test_class*>
.
The list will contain pointers to test_class
objects tightly packed in a std::vector<test_class>
object. We'll call this configuration
compact pointer list
std::list<test_class*>
.
The list will contain pointers to test_class
objects owned by a std::list<test_class>
object. We'll call this configuration disperse pointer list.
Both test_class
and itest_class
are templatized classes that
can be configured with a boolean to increase the size of the object. This way,
the tests can be executed with small and big objects. Here is the first part
of the testing code, which shows the definitions of test_class
and itest_class
classes, and
some other utilities:
//Iteration and element count defines const int NumIter = 100; const int NumElements = 50000; using namespace boost::intrusive; template<bool BigSize> struct filler { int dummy[10]; }; template <> struct filler<false> {}; template<bool BigSize> //The object for non-intrusive containers struct test_class : private filler<BigSize> { int i_; test_class() {} test_class(int i) : i_(i) {} friend bool operator <(const test_class &l, const test_class &r) { return l.i_ < r.i_; } friend bool operator >(const test_class &l, const test_class &r) { return l.i_ > r.i_; } }; template <bool BigSize, link_mode_type LinkMode> struct itest_class //The object for intrusive containers : public list_base_hook<link_mode<LinkMode> >, public test_class<BigSize> { itest_class() {} itest_class(int i) : test_class<BigSize>(i) {} }; template<class FuncObj> //Adapts functors taking values to functors taking pointers struct func_ptr_adaptor : public FuncObj { typedef typename FuncObj::first_argument_type* first_argument_type; typedef typename FuncObj::second_argument_type* second_argument_type; typedef typename FuncObj::result_type result_type; result_type operator()(first_argument_type a, second_argument_type b) const { return FuncObj::operator()(*a, *b); } };
As we can see, test_class
is
a very simple class holding an int
.
itest_class
is just a class
that has a base hook (list_base_hook
)
and also derives from test_class
.
func_ptr_adaptor
is just a
functor adaptor to convert function objects taking test_list
objects to function objects taking pointers to them.
You can find the full test code code in the perf_list.cpp source file.
The first test will measure the benefits we can obtain with intrusive containers
avoiding memory allocations and deallocations. All the objects to be inserted
in intrusive containers are allocated in a single allocation call, whereas
std::list
will need to allocate memory for each
object and deallocate it for every erasure (or container destruction).
Let's compare the code to be executed for each container type for different insertion tests:
std::vector<typename ilist::value_type> objects(NumElements); ilist l; for(int i = 0; i < NumElements; ++i) l.push_back(objects[i]); //Elements are unlinked in ilist's destructor //Elements are destroyed in vector's destructor
For intrusive containers, all the values are created in a vector and after that inserted in the intrusive list.
stdlist l; for(int i = 0; i < NumElements; ++i) l.push_back(typename stdlist::value_type(i)); //Elements unlinked and destroyed in stdlist's destructor
For a standard list, elements are pushed back using push_back().
std::vector<typename stdlist::value_type> objects(NumElements); stdptrlist l; for(int i = 0; i < NumElements; ++i) l.push_back(&objects[i]); //Pointers to elements unlinked and destroyed in stdptrlist's destructor //Elements destroyed in vector's destructor
For a standard compact pointer list, elements are created in a vector and pushed back in the pointer list using push_back().
stdlist objects; stdptrlist l; for(int i = 0; i < NumElements; ++i){ objects.push_back(typename stdlist::value_type(i)); l.push_back(&objects.back()); } //Pointers to elements unlinked and destroyed in stdptrlist's destructor //Elements unlinked and destroyed in stdlist's destructor
For a disperse pointer list, elements are created in a list and pushed back in the pointer list using push_back().
These are the times in microseconds for each case, and the normalized time:
Table 10.2. Back insertion + destruction times for Visual C++ 7.1 / Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
5000 / 22500 |
1 / 1 |
|
7812 / 32187 |
1.56 / 1.43 |
|
10156 / 41562 |
2.03 / 1.84 |
Standard list |
76875 / 97500 |
5.37 / 4.33 |
Standard compact pointer list |
76406 / 86718 |
15.28 / 3.85 |
Standard disperse pointer list |
146562 / 175625 |
29.31 / 7.80 |
Table 10.3. Back insertion + destruction times for GCC 4.1.1 / MinGW over Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
4375 / 22187 |
1 / 1 |
|
7812 / 32812 |
1.78 / 1.47 |
|
10468 / 42031 |
2.39 / 1.89 |
Standard list |
81250 / 98125 |
18.57 / 4.42 |
Standard compact pointer list |
83750 / 94218 |
19.14 / 4.24 |
Standard disperse pointer list |
155625 / 175625 |
35.57 / 7.91 |
Table 10.4. Back insertion + destruction times for GCC 4.1.2 / Linux Kernel 2.6.18 (OpenSuse 10.2)
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
4792 / 20495 |
1 / 1 |
|
7709 / 30803 |
1.60 / 1.5 |
|
10180 / 41183 |
2.12 / 2.0 |
Standard list |
17031 / 32586 |
3.55 / 1.58 |
Standard compact pointer list |
27221 / 34823 |
5.68 / 1.69 |
Standard disperse pointer list |
102272 / 60056 |
21.34 / 2.93 |
The results are logical: intrusive lists just need one allocation. The destruction
time of the normal_link
intrusive
container is trivial (complexity: O(1)
),
whereas safe_link
and auto_unlink
intrusive containers need to
put the hooks of erased values in the default state (complexity: O(NumElements)
). That's why normal_link
intrusive list shines in this test.
Non-intrusive containers need to make many more allocations and that's why
they lag behind. The disperse pointer list
needs to make NumElements*2
allocations,
so the result is not surprising.
The Linux test shows that standard containers perform very well against intrusive containers with big objects. Nearly the same GCC version in MinGW performs worse, so maybe a good memory allocator is the reason for these excellent results.
The next test measures the time needed to complete calls to the member function
reverse()
.
Values (test_class
and itest_class
) and lists are created as explained
in the previous section.
Note that for pointer lists, reverse
does not need to access test_class
values stored in another list or vector, since this function just
needs to adjust internal pointers, so in theory all tested lists need to
perform the same operations.
These are the results:
Table 10.5. Reverse times for Visual C++ 7.1 / Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
2656 / 10625 |
1 / 1.83 |
|
2812 / 10937 |
1.05 / 1.89 |
|
2710 / 10781 |
1.02 / 1.86 |
Standard list |
5781 / 14531 |
2.17 / 2.51 |
Standard compact pointer list |
5781 / 5781 |
2.17 / 1 |
Standard disperse pointer list |
10781 / 15781 |
4.05 / 2.72 |
Table 10.6. Reverse times for GCC 4.1.1 / MinGW over Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
2656 / 10781 |
1 / 2.22 |
|
2656 / 10781 |
1 / 2.22 |
|
2812 / 10781 |
1.02 / 2.22 |
Standard list |
4843 / 12500 |
1.82 / 2.58 |
Standard compact pointer list |
4843 / 4843 |
1.82 / 1 |
Standard disperse pointer list |
9218 / 12968 |
3.47 / 2.67 |
Table 10.7. Reverse times for GCC 4.1.2 / Linux Kernel 2.6.18 (OpenSuse 10.2)
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
2742 / 10847 |
1 / 3.41 |
|
2742 / 10847 |
1 / 3.41 |
|
2742 / 11027 |
1 / 3.47 |
Standard list |
3184 / 10942 |
1.16 / 3.44 |
Standard compact pointer list |
3207 / 3176 |
1.16 / 1 |
Standard disperse pointer list |
5814 / 13381 |
2.12 / 4.21 |
For small objects the results show that the compact storage of values in intrusive containers improve locality and reversing is faster than with standard containers, whose values might be dispersed in memory because each value is independently allocated. Note that the dispersed pointer list (a list of pointers to values allocated in another list) suffers more because nodes of the pointer list might be more dispersed, since allocations from both lists are interleaved in the code:
//Object list (holding `test_class`) stdlist objects; //Pointer list (holding `test_class` pointers) stdptrlist l; for(int i = 0; i < NumElements; ++i){ //Allocation from the object list objects.push_back(stdlist::value_type(i)); //Allocation from the pointer list l.push_back(&objects.back()); }
For big objects the compact pointer list wins because the reversal test doesn't need access to values stored in another container. Since all the allocations for nodes of this pointer list are likely to be close (since there is no other allocation in the process until the pointer list is created) locality is better than with intrusive containers. The dispersed pointer list, as with small values, has poor locality.
The next test measures the time needed to complete calls to the member function
sort(Pred pred)
. Values (test_class
and itest_class
) and lists
are created as explained in the first section. The values will be sorted
in ascending and descending order each iteration. For example, if l
is a list:
for(int i = 0; i < NumIter; ++i){ if(!(i % 2)) l.sort(std::greater<stdlist::value_type>()); else l.sort(std::less<stdlist::value_type>()); }
For a pointer list, the function object will be adapted using func_ptr_adaptor
:
for(int i = 0; i < NumIter; ++i){ if(!(i % 2)) l.sort(func_ptr_adaptor<std::greater<stdlist::value_type> >()); else l.sort(func_ptr_adaptor<std::less<stdlist::value_type> >()); }
Note that for pointer lists, sort
will take a function object that will access test_class
values stored in another list
or vector, so pointer lists will suffer an extra indirection:
they will need to access the test_class
values stored in another container to compare two elements.
These are the results:
Table 10.8. Sort times for Visual C++ 7.1 / Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
16093 / 38906 |
1 / 1 |
|
16093 / 39062 |
1 / 1 |
|
16093 / 38906 |
1 / 1 |
Standard list |
32343 / 56406 |
2.0 / 1.44 |
Standard compact pointer list |
33593 / 46093 |
2.08 / 1.18 |
Standard disperse pointer list |
46875 / 68593 |
2.91 / 1.76 |
Table 10.9. Sort times for GCC 4.1.1 / MinGW over Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
15000 / 39218 |
1 / 1 |
|
15156 / 39531 |
1.01 / 1.01 |
|
15156 / 39531 |
1.01 / 1.01 |
Standard list |
34218 / 56875 |
2.28 / 1.45 |
Standard compact pointer list |
35468 / 49218 |
2.36 / 1.25 |
Standard disperse pointer list |
47656 / 70156 |
3.17 / 1.78 |
Table 10.10. Sort times for GCC 4.1.2 / Linux Kernel 2.6.18 (OpenSuse 10.2)
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
18003 / 40795 |
1 / 1 |
|
18003 / 41017 |
1 / 1 |
|
18230 / 40941 |
1.01 / 1 |
Standard list |
26273 / 49643 |
1.45 / 1.21 |
Standard compact pointer list |
28540 / 43172 |
1.58 / 1.05 |
Standard disperse pointer list |
35077 / 57638 |
1.94 / 1.41 |
The results show that intrusive containers are faster than standard containers. We can see that the pointer list holding pointers to values stored in a vector is quite fast, so the extra indirection that is needed to access the value is minimized because all the values are tightly stored, improving caching. The disperse list, on the other hand, is slower because the indirection to access values stored in the object list is more expensive than accessing values stored in a vector.
The next test measures the time needed to iterate through all the elements
of a list, and increment the value of the internal i_
member:
stdlist::iterator it(l.begin()), end(l.end()); for(; it != end; ++it) ++(it->i_);
Values (test_class
and itest_class
) and lists are created as explained
in the first section. Note that for pointer lists, the iteration will suffer
an extra indirection: they will need to access the test_class
values stored in another container:
stdptrlist::iterator it(l.begin()), end(l.end()); for(; it != end; ++it) ++((*it)->i_);
These are the results:
Table 10.11. Write access times for Visual C++ 7.1 / Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
2031 / 8125 |
1 / 1 |
|
2031 / 8281 |
1 / 1.01 |
|
2031 / 8281 |
1 / 1.01 |
Standard list |
4218 / 10000 |
2.07 / 1.23 |
Standard compact pointer list |
4062 / 8437 |
2.0 / 1.03 |
Standard disperse pointer list |
8593 / 13125 |
4.23 / 1.61 |
Table 10.12. Write access times for GCC 4.1.1 / MinGW over Windows XP
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
2343 / 8281 |
1 / 1 |
|
2500 / 8281 |
1.06 / 1 |
|
2500 / 8281 |
1.06 / 1 |
Standard list |
4218 / 10781 |
1.8 / 1.3 |
Standard compact pointer list |
3906 / 8281 |
1.66 / 1 |
Standard disperse pointer list |
8281 / 13750 |
3.53 / 1.66 |
Table 10.13. Write access times for GCC 4.1.2 / Linux Kernel 2.6.18 (OpenSuse 10.2)
Container |
Time in us/iteration (small object / big object) |
Normalized time (small object / big object) |
---|---|---|
|
2286 / 8468 |
1 / 1.1 |
|
2381 / 8412 |
1.04 / 1.09 |
|
2301 / 8437 |
1.01 / 1.1 |
Standard list |
3044 / 9061 |
1.33 / 1.18 |
Standard compact pointer list |
2755 / 7660 |
1.20 / 1 |
Standard disperse pointer list |
6118 / 12453 |
2.67 / 1.62 |
As with the read access test, the results show that intrusive containers outperform all other containers if the values are tightly packed in a vector. The disperse list is again the slowest.
Intrusive containers can offer performance benefits that cannot be achieved
with equivalent non-intrusive containers. Memory locality improvements are
noticeable when the objects to be inserted are small. Minimizing memory allocation/deallocation
calls is also an important factor and intrusive containers make this simple
if the user allocates all the objects to be inserted in intrusive containers
in containers like std::vector
or std::deque
.