Crystalline Reclamation
Publications
- Brief Announcement: Crystalline: Fast and Memory Efficient Wait-Free Reclamation. Ruslan Nikolaev and Binoy Ravindran. In Proceedings of the 35th International Symposium on Distributed Computing (DISC'21). Freiburg, Germany.
Source Code
This code further extends the existent Interval-Based-Reclamation benchmark. Crystalline's code is directly embedded into the benchmark. Hyaline's code is in a separate 'hyaline' directory.
Building
To build this benchmark suite, we use clang 11.0. You can download it from (http://releases.llvm.org/download.html). You can use Ubuntu 16.04 pre-built binaries on Ubuntu 18.04 LTS:
Please extract it to /opt/llvm/ (e.g., /opt/llvm/bin/clang for clang). If extracted to a different directory, then change clang and clang++ paths in the Makefile.
You also need to install hwloc and jemalloc. If running Ubuntu 18.04, you can type:
sudo apt-get install build-essential libjemalloc-dev libhwloc-dev libc6-dev libc-dev make python
To compile the benchmark with Crystalline:
-
Go to 'benchmark' (This benchmark is already modified to use Crystalline.)
-
Run 'make'
See the original IBR's instructions in the 'benchmark' directory.
Unreclaimed Objects
We had to change the default method of counting uncreclaimed objects in the benchmark, as the original approach would not work as is with Hyaline or Crystalline due to the global retirement of objects.
However, measurements become more expensive and skew throughput for some tests. For this reason, we introduced an extra '-c' flag. By default, we do not count (properly) the number of unreclaimed objects. If the -c flag is specified, we count the number of objects (potentially skewing throughput). To count the number of objects, you have to run a separate test with the -c flag.
Usage Example
A sample command (not counting uncreclaimed objects):
./bin/main -i 1 -m 3 -v -r 1 -o hashmap_result.csv -t 4 -d tracker=HR