For details, please refer to our VLDB paper "Plush: A Write-Optimized Persistent Log-Structured Hash-Table".
Reference forthcoming.
Copyright (c) 2022 TUM. All rights reserved.
AVX-512 support, a mounted PMem partition in fsdax mode, g++ >= 11. Some of the other data structures also require pmdk.
Note: While we tried to make the data structure as generic as possible,
some constants are specific to our personal pmem server setup.
You might need to change a few paths, e.g., the path to the pmem partition in src/benchmarking/PiBenchWrapper.hpp.
Plush itself is implemented in src/hashtable/hashtable.cpp and src/hashtable/hashtable.h.
The target hashtable builds a dynamic library which can be included in any other program.
This library also already includes PiBench bindings and can thus be called natively
by the PiBench benchmarking framework by Lersch et al.
We include patched, PiBench-enabled versions of the data structures we compare against as git submodules.
To populate them, run git submodule update --init --recursive.
We have added the raw benchmarking results used to generate the plots paper in the scripts/benchmarks subdirectory.
If you want to verify the results yourself, you can use the shell scripts in the scripts directory.
You can then use the R scripts in the scripts/R subdirectory to generate the figures as pdf and tex files by executing
Rscript filename.R.
To get Plush's performance without storing filters in DRAM, set the constant MAX_DRAM_FILTER_LEVEL in hashtable.h to -1.
We have also added two small example programs to show how we intend Plush to be used. You can find them in the examples directory.
They're also cmake build targets.
To run them, please modify the paths according to your local machine's PMem mount location.
example.cpp shows all features of Plush in a minimal example.
gutenberg.cpp demonstrates Plush's support for variable sized records by loading and storing all ebooks
of Project Gutenberg. Note that this is quite slow as we only use a single thread and most time is spend
loading the books (naively) into main memory.
You can play around by changing all the constants in hashtable.h, but please note that not all constants can currently
be choosen freely. However, if initialization of Plush seems quite slow, you might want to play around with
PAYLOAD_LOG_NUM_BITS
PAYLOAD_CHUNK_NUM_BITS
CHUNKS_PER_PAYLOAD_LOG
PAYLOAD_CHUNK_SIZE