These are the programs used to produce the results found in N4468, "On Quantifying Allocation Strategies". Some of these programs have had bugs fixed since the paper was produced, so may not produce identical results.
These programs depend upon:
- Clang version 3.6 or later, or gcc version 5.1 or later, built with support for link-time-optimization (LTO, see FAQ)
- For clang, libc++ version 3.6, or later
- For gcc 5.1, libstdc++ patched according to: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66055
- BDE Tools, which contains a custom build system based on waf
- If building for LTO, /usr/bin/ld indicating ld.gold
A suitably-configured system can be built quickly using Docker; see the 'Docker' section below for instructions.
This repository contains a snapshot of the BDE library patched to make use of various C++14 features required for the benchmark programs. The patches are included separately here in the [bde-patches-minimal] and [bde-patches-opt] files.
To configure,
$ vi Makefile # configure per instructions
To build,
$ make build
To run benchmarks,
$ make run
Other targets of interest:
bde growth locality zation tention growth-orig
run-growth run-locality run-zation run-tention
clean
| progran | section | what |
|---|---|---|
| growth.cc | section 7 | Creating/destroying isolated basic data structures |
| locality.cc | section 8 | Variation in Locality (long running) |
| zation.cc | section 9 | Variation in Utilization |
| tention.cc | section 10 | Variation in Contention |
Other files:
| file | what |
|---|---|
| readme-growth.txt | instructions to produce CSV of tables in the paper |
| test-growth | scripts to run the benchmarks as published |
| test-locality | |
| test-zation | |
| test-tention | |
| bde-patches-minimal | snapshot of patches to bde that this depends on |
| bde-patches-opt | snapshot of an optimization for (multi-)pool |
Two Dockerfiles have been provided which can be used to construct images with all necessary tools and configuration for building these programs. One file builds a Debian 'testing' (Stretch) based image, and the other builds an Ubuntu 15.04 (Vivid) based image.
To use the Debian Stretch Dockerfile, follow these steps:
# Build an image called 'debian-stretch-clang'
docker build -f ./docker/debian-stretch-clang --rm -t debian-stretch-clang ./docker
# Build the BDE library and benchmarks using the image
docker run --rm -v $(git rev-parse --show-toplevel):/src -w /src/benchmarks/allocators debian-stretch-clang make build
# Run the benchmarks
docker run --rm -v $(git rev-parse --show-toplevel):/src -w /src/benchmarks/allocators debian-stretch-clang make run
To use the Ubuntu Vivid Dockerfile, follow these steps:
# Build an image called 'ubuntu-vivid-clang'
docker build -f ./docker/ubuntu-vivid-clang --rm -t ubuntu-vivid-clang ./docker
# Build the BDE library and benchmarks using the image
docker run --rm -v $(git rev-parse --show-toplevel):/src -w /src/benchmarks/allocators ubuntu-vivid-clang make build
# Run the benchmarks
docker run --rm -v $(git rev-parse --show-toplevel):/src -w /src/benchmarks/allocators ubuntu-vivid-clang make run
Q1: What about tcmalloc?
A1: In all our tests, tcmalloc was substantially slower than the default new/delete on the target platform.
Q2: Why do these depend on recent clang++/libc++ or g++/libstdc++?
A2: The tests, particularly growth.cc and locality.cc, depend on features of C++14: specifically, the containers' comprehensive observance of allocator- traits requirements to direct their memory management.
Q3: How can I build these with a non-LTO-enabled toolchain?
A3: Comment out the LTO variable value in Makefile.