Boost C++ Libraries

...one of the most highly regarded and expertly designed C++ library projects in the world. Herb Sutter and Andrei Alexandrescu, C++ Coding Standards

This is the documentation for an old version of Boost. Click here to view this page for the latest version.

Boost.Flyweight Performance




Contents

Introduction

We show how to estimate the memory reduction obtained by the usage of Boost.Flyweight in a particular scenario and study the impact on the execution time for the different functional areas of flyweight. Some experimental results are provided.

Memory consumption

As we saw in the tutorial rationale, the flyweight pattern is based on two types of objects:

The overall memory consumption is then a function of the size of the flyweight objects, the size of the entry objects and the degree of value redundancy.

Flyweight size

The only data member of a flyweight object is a so-called handle, an opaque object of small size provided by the internal flyweight factory to refer to the entries it stores. For the default hashed_factory, this handle is merely a pointer, so sizeof(flyweight<T>)=sizeof(void*), 4 bytes in typical 32-bit architectures. For other types of factories, the handle is an iterator to an internal container used in the implementation of the factory: again, its size is typically that of a pointer.

Entry size

The entries stored in the factory associated to flyweight<T,...> need not only hold a value of T, but also contain additional information related to the internal implementation of flyweight<T,...>:

entry = sizeof(T) + overhead.

For the current implementation of Boost.Flyweight, the following aspects contribute to overhead:

The table summarizes the separate contributions to overhead introduced by the different components taking part of the definition of a flyweight instantiation. Values are given in words, i.e. the size of a pointer, which is 4 bytes in a typical 32-bit architecture. Alignment may introduce additional overhead.

Entry overhead of the components of Boost.Flyweight.
component overhead (words)
  key_value     with key extractor     1(1)  
  without key extractor     1 + sizeof(Key)  
  factory     hashed_factory     ~2.5  
  set_factory     4(2)  
  assoc_container_factory     depends on the container used  
  tracking mechanism     refcounted     2(3)  
  no_tracking     0  
(1) Assuming that sizeof(Key)<=sizeof(Value).
(2) For some implementations of std::set this overhead reduces to 3.
(3) In some platforms this value can be 3.

For instance, for the default configuration parameters of flyweight, overhead is typically 2.5(hashed_factory) + 2(refcounted) = 4.5 words.

Overall memory consumption

Consider a scenario where there are N different objects of type T jointly taking M different values. The objects consume then S = N·T bytes, where T is defined as the average size of T (sizeof(T) plus dynamic memory allocated by T objects). If we now replace T by some instantiation flyweight<T,...>, the resulting memory consumption will be

SF = N·P + M·(T + overhead),

where P is sizeof(flyweight<T,...>), typically equal to sizeof(void*), as seen before. The ratio SF / S is then

SF / S = (P / T)+ (M / N)(1 + overhead / T).

SF / S tends to its minimum, P / T, as M / N tends to 0, i.e. when the degree of value redundancy among T objects grows higher. On the other hand, the worst possible case SF / S = 1 + (P + overhead) / T happens when M / N = 1, that is, if there is no value redundancy at all; in this situation there is no point in applying the flyweight pattern in the first place.

relative memory consumption of Boost.Flyweight as a function of value diversity
Fig. 1: Relative memory consumption of Boost.Flyweight as a function of value diversity.

Time efficiency

The introduction of the flyweight pattern involves an extra level of indirection that, in general, results in some execution overhead when accessing the values. On the other hand, manipulation of flyweight objects is considerably faster than moving around the heavy values they stand for. We analyze qualitatively the execution overheads or improvements associated to the different usage contexts of Boost.Flyweight.

Initialization

As compared with the initialization an object of type T, constructing a flyweight<T> performs important extra work like looking up the value in the flyweight factory and inserting it if it is not present. So, construction of flyweights (other than copy construction, which is cheap), is expected to be noticeably slower than the construction of the underlying type T. Much of the time spent at constructing the associated T value proper can be saved, however, by using key-value flyweights.

Assignment

Assignment of flyweight objects is extremely fast, as it only involves assigning an internal handle type used to refer to the shared value. Moreover, assignment of flyweight objects never throws. Assignment time is influenced by the type of tracking policy used; in this regard, no_tracking is the fastest option.

Equality comparison

Comparing two flyweight objects for equality reduces to checking that the addresses of the values they are associated to are equal; in general, this operation is much faster than comparing the underlying values. This aspect is of particular relevance when the flyweight objects stand for complex values like those arising in the application of the composite pattern.

Value access

The conversion from flyweight<T> to const T& relies on a level of indirection relating the flyweight objects to the values they are associated to; so, value access is expected to be slower when using Boost.Flyweight as compared to using the associated values directly. This overhead, however, can be masked by an indirect improvement resulting from locality and cache effects: as the set of different T values handled by an instantiation of flyweight<T> is generally much smaller than the equivalent family of T objects when Boost.Flyweight is not used, active values can fit better into the processor cache.

Experimental results

A profiling program was devised to test the space and time efficiency of different instantiations of flyweight against a base situation not using Boost.Flyweight. The profiled scenarios are:

  1. std::string.
  2. flyweight<std::string> with default configuration aspects (hashed_factory, refcounted tracking, simple_locking).
  3. flyweight<std::string,no_tracking>.
  4. flyweight<std::string,set_factory>.
  5. flyweight<std::string,set_factory,no_tracking>.

Actually the types tested are not exactly those listed above, but instrumented versions that keep track of the allocated memory for profiling purposes. The program parses a text file into an array of words and then perform various manipulations involving the different context usages of Boost.Flyweight discussed previously. As our text file we have used the plain text version of Project Gutenberg edition of Don Quijote (2.04 MB).

Microsoft Visual C++ 8.0

The program was built with default release settings and _SECURE_SCL=0. Tests were run under Windows XP in a machine equipped with an Intel Core 2 Duo T5500 processor and 1 GB of RAM.

Memory

memory consumption (MB), MSVC++ 8.0
Fig. 2: Memory consumption, MSVC++ 8.0. Values in MB.

The results show the memory consumption figures for the different profiled scenarios. The standard library implementation of MSVC++ 8.0 features the so-called small buffer optimization for strings, by which std::string objects hold a small buffer that can be used when the string is short, thus avoding dynamic allocations. This results in sizeof(std::string) being quite high, 28 bytes. In our particular test strings are almost always held in the small buffer, so the minimum SF / S achievable is 4/28 = 14.3%, which is quite close to the experimental results, given that the memory devoted to storage of shared values is residual (around 3% of total memory) due to the high word redundancy of the text source.

Execution time

execution time (s), MSVC++ 8.0
Fig. 3: Execution time, MSVC++ 8.0. Values in seconds.

The figure displays execution times for the profiled scenarios in different usage contexts. In accordance with our previous qualitative analysis, initialization of flyweights carries an important overhead with respect to the base case scenario (between 20% and 40% of additional execution time), while the other usage contexts (assignment, equality comparison and value access) have performance gains, with speedup factors of more than 10 in some cases. The use of a refcounted tracking policy introduces penalties with respect to no_tracking in initialization and assignment, but has no effect in equality comparison and value access.

GNU GCC 3.4.4

The Cygwin/MinGW version of the compiler was used, with command options -ftemplate-depth-128 -O3 -finline-functions -DNDEBUG. Tests were run under a Cygwin terminal in the same machine as before.

Memory

memory consumption (MB), GCC 3.4.4
Fig. 4: Memory consumption, GCC 3.4.4. Values in MB.

The standard library used by GCC 3.4.4 implements std::string using copy-on-write optimization techniques, which leads to very small value redundancy for some usage patterns. This explains why the memory reduction achieved by Boost.Flyweight is so poor in this case. Other contexts where assignment is much less used than direct construction will favor Boost.Flyweight over plain copy-on-write std::strings.

Execution time

execution time (s), GCC 3.4.4
Fig. 5: Execution time, GCC 3.4.4. Values in seconds.

Relative performance figures are similar to those obtained for MSVC++ 8.0, although some of the speedups achieved by Boost.Flyweight are higher here (×25 in equality comparison and up to ×100 in assignment when no_tracking is in effect).

Conclusions

The introduction of Boost.Flyweight in application scenarios with very high value redundancy yields important reductions in memory consumption: this is especially relevant when data volume approaches the limits of physical memory in the machine, since Boost.Flyweight can avoid virtual memory thrashing thus making the application viable. We have shown how to estimate the achievable reduction in memory consumption from some basic value statistics and knowledge of the flyweight configuration aspects being used.

Boost.Flyweight can also accelerate execution times in areas other than object initialization, due to the fastest manipulation of small flyweight objects and to locality and cache effects arising from the drastic reduction of the set of allocated values.





Revised June 22nd 2009

© Copyright 2006-2009 Joaquín M López Muñoz. Distributed under the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)