Click here to Skip to main content
15,889,216 members
Articles / Programming Languages / C++
Article

A Custom Block Allocator for Speeding Up VC++ STL

Rate me:
Please Sign up or sign in to vote.
4.63/5 (20 votes)
30 Oct 2006CPOL 237.5K   1.5K   59   63
A block allocator for use with STL containers that greatly improves speed in programs doing massive data insertions and extractions.

Introduction

block_allocator is a custom STL allocator for use with STL as implemented in Microsoft VC++. Rather than doing allocations on a per-node basis, block_allocator allocates memory in fixed sized chunks, and delivers portions of these chunks as requested. Typical speed improvements of 40% have been obtained with respect to the default allocator. The size of the chunks, set by the user, should not be too little (reduced speed improvements) nor too large (memory wasted). Experiment and see what sizes fit best to your application.

block_allocator can substitute for the default allocator in the following containers:

  • list,
  • set,
  • multiset,
  • map,
  • multimap,
and WON'T work with other containers such as vector or queue. Note however that vector and queue already perform allocation in chunks. The usage of block_allocator is fairly simple, for instance:
// block allocated list of ints with chunks of 1024 elements
std::list<int,block_allocator<int,1024> > l;
Normal containers and block allocated containers can coexist without problems.

Compatibility mode with MSVC++ 6.0/7.0

Due to limitations of the standard library provided with these compilers, the mode of usage explained above does not work here. To circumvent this problem one must proceed as follows: For each of the containers supported, there's an associated block allocated container derived from it thru use of block_allocator. You have to define an activating macro for each container to be defined prior to the inclusion of blockallocator.h:

  • list -> block_allocated_list (macro DEFINE_BLOCK_ALLOCATED_LIST),
  • set -> block_allocated_set (macro DEFINE_BLOCK_ALLOCATED_SET),
  • multiset -> block_allocated_multiset (macro DEFINE_BLOCK_ALLOCATED_MULTISET),
  • map -> block_allocated_map (macro DEFINE_BLOCK_ALLOCATED_MAP),
  • multimap -> block_allocated_multimap (macro DEFINE_BLOCK_ALLOCATED_MULTIMAP),

To use block allocation based STL in your application, define the corresponding activating macro, include blockallocator.h and then change your declarations as follows:

  • list<type> -> block_allocated_list<type,chunk_size>
  • set<key> -> block_allocated_set<key,chunk_size>
  • multiset<key> -> block_allocated_multiset<key,chunk_size>
  • map<key,type> -> block_allocated_map<key,type,chunk_size>
  • multimap<key,type> -> block_allocated_multimap<key,type,chunk_size>

where chunk_size is the size of the chunks. You can enter too the other optional template parameters (see MSVC++ STL docs for more info).

The MSVC++ 6.0/7.0 compatibility mode can also be used in MSVC++ 7.1, so you need not modify your block_allocator-related code when porting legacy code to 7.1.

Multithreading issues

Each block allocated container instance uses its own block_allocator, so no multithreading problems should arise as long as your program conveniently protects their containers for concurrent access (or if no two threads access the same container instance). This is the same scenario posed by regular STL classes (remember operations on containers are not guarded by CRITICAL_SECTIONs or anything similar), so the moral of it all is: If your program was multithread safe without block_allocator, it'll continue to be with it.

Version history

  • 29th Feb, 2000 - 1.1
    • Initial release in CodeProject.
  • 22nd Mar, 2001 - 1.2
    • Included definitions for operator== and operator!=. The lack of these caused linking errors when invoking list::swap() and similar methods. The funny thing about it is that no one ever reported this seemingly important bug, so either swap() is not that much used or not that many people use block_allocator!
  • 25th Oct, 2006 - 1.3
    • block_allocator now works with MSVC++ 7.1 and 8.0. Thanks to James May for helping with testing this new version of the code.
  • 30th Oct, 2006 - 1.4
    • Fixed some typedefs incorrectly made private in block_allocated_list, block_allocated_set, etc.

License

This article, along with any associated source code and files, is licensed under The Code Project Open License (CPOL)


Written By
Spain Spain
This member has not yet provided a Biography. Assume it's interesting and varied, and probably something to do with programming.

Comments and Discussions

 
AnswerRe: Will this decrease size? Pin
Joaquín M López Muñoz19-Feb-02 20:10
Joaquín M López Muñoz19-Feb-02 20:10 
Generalusing with dinkumware library Pin
User 988514-Feb-02 10:26
User 988514-Feb-02 10:26 
GeneralRe: using with dinkumware library Pin
Joaquín M López Muñoz14-Feb-02 20:41
Joaquín M López Muñoz14-Feb-02 20:41 
QuestionHow can such an allocator be used in a vector? Pin
User 98855-Feb-02 5:32
User 98855-Feb-02 5:32 
AnswerRe: How can such an allocator be used in a vector? Pin
Joaquín M López Muñoz5-Feb-02 6:01
Joaquín M López Muñoz5-Feb-02 6:01 
Generalblockallocator bug Pin
6-Mar-01 10:20
suss6-Mar-01 10:20 
GeneralRe: blockallocator bug Pin
Joaquín M López Muñoz7-Mar-01 21:58
Joaquín M López Muñoz7-Mar-01 21:58 
Generalblockallocator bug: closer inspection and proposed workaround Pin
Joaquín M López Muñoz2-Oct-01 0:12
Joaquín M López Muñoz2-Oct-01 0:12 
This may come a little late months after your post, but now that I had some time to release v1.2 of block_allocator I feel like I can clarify this issue.
The bug you reported is caused (as you pointed out) to the deallocation of an static struct called _Nil by a different block_allocator instance than the one that allocated it. In my opinion, it is the VC++ STL implementation the one to blame for this bug, as it fails to check whether allocators are interchangeable before trying deallocation (however, this check is systematically performed at all other places where it matters).
One simple workaround for this problem is having a static container of the offending type so that both allocation and deallocation of _Nil are carried out by it regardless of the various containers of the same type being constructed and destructed during the life of the program. The following sample code shows the technique:
#define DEFINE_BLOCK_ALLOCATED_MAP
#include "blockallocator.h"

#include <crtdbg.h>

using namespace std;

typedef block_allocated_map<int,int,1024> bmap;

//static bmap last_to_abandon; /* comment out/in to fix/show the problem */

int main(void)
{
  _CrtSetDbgFlag(_CRTDBG_DELAY_FREE_MEM_DF|_CrtSetDbgFlag(_CRTDBG_REPORT_FLAG));

  bmap *m1=new bmap;
  bmap *m2=new bmap;

  delete m1;
  delete m2;

  _CrtCheckMemory();
  
  return 0;
}
As for the multithreading issues I mentioned in my first reply to your post, a closer reevaluation of the library shows it is as multithread safe as it can be (more info in the updated article above).

Hope this helps.

Joaquín M López Muñoz
Telefónica, Investigación y Desarrollo
QuestionIs the allocator thread-safe? Pin
29-Dec-00 5:52
suss29-Dec-00 5:52 
AnswerRe: Is the allocator thread-safe? Pin
Joaquín M López Muñoz2-Oct-01 0:20
Joaquín M López Muñoz2-Oct-01 0:20 
General12% map speed improvement achieved by using this approach Pin
26-Nov-00 5:18
suss26-Nov-00 5:18 
QuestionWhere is the speedup? Pin
Jim16-May-00 2:12
Jim16-May-00 2:12 
AnswerRe: Where is the speedup? Pin
Joaquín M López Muñoz16-May-00 3:18
Joaquín M López Muñoz16-May-00 3:18 
AnswerRe: Where is the speedup? Pin
Joaquín M López Muñoz8-Mar-01 21:23
Joaquín M López Muñoz8-Mar-01 21:23 
GeneralRe: Where is the speedup? Pin
19-Apr-01 14:19
suss19-Apr-01 14:19 
GeneralRe: Where is the speedup? Pin
6-Sep-01 6:44
suss6-Sep-01 6:44 
GeneralRe: Where is the speedup? Pin
14-Sep-01 9:03
suss14-Sep-01 9:03 

General General    News News    Suggestion Suggestion    Question Question    Bug Bug    Answer Answer    Joke Joke    Praise Praise    Rant Rant    Admin Admin   

Use Ctrl+Left/Right to switch messages, Ctrl+Up/Down to switch threads, Ctrl+Shift+Left/Right to switch pages.