https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113504
Bug ID: 113504
Summary: High memory usage for parallel `std::sort`
Product: gcc
Version: 12.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: ruben.laso at tuwien dot ac.at
Target Milestone: ---
The memory usage of parallel `std::sort` is very high compared to the
sequential version and even other parallel implementations.
The attached code is a simple test case to compare the memory usage of parallel
`std::sort`, `tbb::parallel_sort` and sequential `std::sort`.
The test case has been replicated in several systems with versions of GCC 10,
11 and 12.
An example of the results (and max. memory usage according to `/usr/bin/time`)
is shown in the following table:
| Executable | Size | Time | Max Resident Memory |
| ------------------ | ----------- | -------- | ------------------- |
| ./pstl_sort.out | 33554432 | 0:00.23 | 423776k |
| ./tbb_sort.out | 33554432 | 0:00.44 | 143952k |
| ./seq_sort.out | 33554432 | 0:03.32 | 134836k |
| ./pstl_sort.out | 1073741824 | 0:05.68 | 14236656k |
| ./tbb_sort.out | 1073741824 | 0:13.02 | 4207680k |
| ./seq_sort.out | 1073741824 | 2:07.38 | 4198124k |
In the example, the parallel `std::sort` (pstl_sort) uses ~3 times more memory
than the `tbb::parallel_sort` (tbb_sort) and the sequential `std::sort`
(seq_sort).
It also runs faster, though.
System specs in the example:
CPU: AMD EPYC 7551
RAM: 256 GB DDR4
OS: Debian 10.10
Compilation with:
g++ -std=c++17 -O3 -pedantic -Wall -Wextra -Werror -o pstl_sort.out main.cpp
-ltbb -DPSTL_SORT
g++ -std=c++17 -O3 -pedantic -Wall -Wextra -Werror -o tbb_sort.out main.cpp
-ltbb -DTBB_SORT
g++ -std=c++17 -O3 -pedantic -Wall -Wextra -Werror -o seq_sort.out main.cpp
-ltbb
Did I miss something in the code?
Is that high memory usage a deliberate trade-off for performance?
Is the algorithm still in development to improve memory usage?