RECIPE proposes a principled approach for converting concurrent indexes built for DRAM into crash-consistent indexes for persistent memory. This repository includes the implementations of the index structures for persistent memory converted from the existing concurrent DRAM indexes by following RECIPE. For performance evaluations, this repository also provides the microbenchmarks for index structures based on YCSB. This repository contains all the information needed to reproduce the main results from our paper.
Please cite the following paper if you use the RECIPE approach or RECIPE-converted indexes:
RECIPE : Converting Concurrent DRAM Indexes to Persistent-Memory Indexes. Se Kwon Lee, Jayashree Mohan, Sanidhya Kashyap, Taesoo Kim, Vijay Chidambaram. Proceedings of the The 27th ACM Symposium on Operating Systems Principles (SOSP 19). Paper PDF. Extended version(arXiv). Bibtex.
@InProceedings{LeeEtAl19-Recipe,
title = "{RECIPE: Converting Concurrent DRAM Indexes to Persistent-Memory Indexes}",
author = "Se Kwon Lee and Jayashree Mohan and Sanidhya Kashyap and Taesoo Kim and Vijay Chidambaram",
booktitle = "Proceedings of the 27th ACM Symposium on Operating
Systems Principles (SOSP '19)",
month = "October",
year = "2019",
address = "Ontario, Canada",
}
- P-CLHT has been used to build DINOMO, a key-value store for disaggregated persistent memory.
The following improvements are made to the codebase after the SOSP paper.
P-CLHT/contains the source code for P-CLHT. It is converted from Cache-Line Hash Table to be persistent. The original source code and paper can be found in code and paper.P-HOT/contains the source code for P-HOT. It is converted from Height Optimized Trie to be persistent. The original source code and paper can be found in code and paper.P-BwTree/contains the source code for P-BwTree. It is converted from an open sourced implementation of BwTree for persistent memory. The original source code and paper can be found in code and paper.P-ART/contains the source code for P-ART. It is converted for persistent memory from Adaptive Radix Tree using ROWEX for concurrency. The original source code and paper can be found in code and paper.P-Masstree/contains the source code for P-Masstree. It is converted from Masstree to be persistent and is custumized for the compact version. The original source code and paper can be found in code and paper.index-microbench/contains the benchmark framework to generate YCSB workloads. The original source code can be found in code.
P-CLHTis a good fit for applications requiring high-performance point queries.P-HOTis a good fit for applications with read-dominated workloads.P-BwTreeprovides well-balanced performance for insertion, lookup, and range scan operations for applications using integer keys.P-ARTis suitable for applications with insertion-dominated workloads and a small number of range queries.P-Masstreeprovides well-balanced performance for insertion, lookup, and range scan operations for applications using either integer or string keys.
Apart from benchmark code with ycsb.cpp, we provide simple example codes (P-*/example.cpp for each RECIPE index)
to help developers who want to apply RECIPE indexes into their own project to easily identify how to use each index's APIs.
These example source codes run insert and lookup operations with custom integer keys. For more details of usage for each index,
please refer to P-*/README.md in each index's directory and ycsb.cpp as well.
The RECIPE data structures in the master branch use a volatile memory allocator (libvmmalloc) so that RECIPE can be compared in an apples-to-apples manner with prior work like FAST&FAIR and CCEH, which also use volatile allocators (and thereby do not provide crash consistency). Thus, if you use RECIPE data structures from the master branch on PM, metadata related to memory allocator will not have crash consistency.
The current volatile allocator must be replaced with persistent memory allocator to ensure crash consistency of memory allocator and to prevent permanent memory leaks. Especially, we recommend post-crash garbage collection rather than logging-based approaches to solve permanent memory leaks since logging-based approaches should constantly consume costs for recording logs during normal runtime (We already described it through our SOSP publication). We are currently exploring various post-crash garbage collection techniques ([1], [2], [3], [4], [5]) to apply them for RECIPE data structures.
As a first step, we are working on replacing current volatile allocator with PMDK library and on
solving permanent memory leaks using the functions provided by it [5]. Please check out
the pmdk branch for the updates of this work as well as these details.
- Ubuntu 18.04.1 LTS
- At least 32GB DRAM
- x86-64 CPU supporting at least 16 threads
- P-HOT: x86-64 CPU supporting at least the AVX-2 and BMI-2 instruction sets (Haswell and newer)
- Compile: cmake, g++-7, gcc-7, c++17
$ sudo apt-get install build-essential cmake libboost-all-dev libpapi-dev default-jdk
$ sudo apt-get install libtbb-dev libjemalloc-dev
$ cd ./index-microbench
$ curl -O --location https://github.com/brianfrankcooper/YCSB/releases/download/0.11.0/ycsb-0.11.0.tar.gz
$ tar xfvz ycsb-0.11.0.tar.gz
$ mv ycsb-0.11.0 YCSB
Configure the options of each workloads (a, b, c, e), would only need to change $recordcount and $operationcount.
$ vi ./index-microbench/workload_spec/<workloada or workloadb or workloadc or workloade>
Select which workloads to be generated. Default configuration will generate all workloads (a, b, c, e). Change the code line for WORKLOAD_TYPE in <a b c e>; do, depending on which workload you want to generate.
$ vi ./index-microbench/generate_all_workloads.sh
Generate the workloads. This will generate both random integer keys and string ycsb keys with the specified key distribution.
$ cd ./index-microbench/
$ mkdir workloads
$ bash generate_all_workloads.sh
Change LOAD_SIZE and RUN_SIZE variables to be same with the generated workload size, which are hard-coded in ycsb.cpp (Default is 64000000).
$ vi ycsb.cpp
For running the indexes on Intel Optane DC Persistent Memory, we will use libvmmalloc to transparently converts all dynamic memory allocations into Persistent Memory allocations, mapped by pmem.
$ sudo mkfs.ext4 -b 4096 -E stride=512 -F /dev/pmem0
$ sudo mount -o dax /dev/pmem0 /mnt/pmem
Install PMDK
$ git clone https://github.com/pmem/pmdk.git $ cd pmdk $ git checkout tags/1.6 $ make -j $ cd ..
Configuration for libvmmalloc
- LD_PRELOAD=path
Specifies a path to libvmmalloc.so.1. The default indicates the path to libvmmalloc.so.1 that is built from the instructions installing PMDK above.
- VMMALLOC_POOR_DIR=path
Specifies a path to the directory where the memory pool file should be created. The directory must exist and be writable.
- VMMALLOC_POOL_SIZE=len
Defines the desired size (in bytes) of the memory pool file.
$ vi ./scripts/set_vmmalloc.sh
Please change below configurations to fit for your environment.
export VMMALLOC_POOL_SIZE=$((64*1024*1024*1024))
export VMMALLOC_POOL_DIR="/mnt/pmem"
Build all
$ mkdir build $ cd build $ cmake .. $ make
Run
$ cd ${project root directory}
$ ./build/ycsb art a randint uniform 4
Usage: ./ycsb [index type] [ycsb workload type] [key distribution] [access pattern] [number of threads]
1. index type: art hot bwtree masstree clht
fastfair levelhash cceh
2. ycsb workload type: a, b, c, e
3. key distribution: randint, string
4. access pattern: uniform, zipfian
5. number of threads (integer)
Run
$ cd ${project root directory}
$ sudo su
# source ./scripts/set_vmmalloc.sh
# LD_PRELOAD="./pmdk/src/nondebug/libvmmalloc.so.1" ./build/ycsb art a randint uniform 4
# source ./scripts/unset_vmmalloc.sh
For artifact evaluation, we will evaluate again the performance of the index structures presented in the paper by using YCSB benchmark. The index structures tested for artifact evaluation include P-CLHT P-ART, P-HOT, P-Masstree, P-Bwtree, FAST&FAIR, WOART, CCEH, and Level hashing. The evaluation results will be stored in ./results directory as csv files. Please make sure to check the contents at least by checklists subsection in Benchmark details section below, before beginning artifact evaluation. Note that the evaluations re-generated for artifact evaluation will be based on DRAM because Optane DC persistent memory machine used for the evaluations presented in the paper has the hard access limitation from external users. For more detail, please refer to experiments.md.
RECIPE has been awarded three badges: Artifact Available, Artifact Functional, and Results Reproduced.
[1] Kumud Bhandari, et al. Makalu: Fast Recoverable Allocation of Non-volatile Memory, OOPSLA'16.
[2] Nachshon Cohen, et al. Object-Oriented Recovery for Non-volatile Memory, OOPSLA'18.
[3] Tudor David, et al. Log-Free Concurrent Data Structures, ATC'18.
[4] Wentao Cai, et al. Understanding and optimizing persistent memory allocation, ISMM'20.
[5] Eduardo B., Code Sample: Find Your Leaked Persistent Memory Objects Using the Persistent Memory Development Kit (PMDK).
The licence for most of the P-* family of persistent indexes is Apache License (https://www.apache.org/licenses/LICENSE-2.0). This is consistent with the most of the indexes we build on, with the exception of CLHT and HOT, which uses the MIT and ISC License respectively. Accordingly, P-CLHT is under the MIT license (https://opensource.org/licenses/MIT). P-HOT is under the ISC license (https://opensource.org/licenses/ISC).
Copyright for RECIPE indexes is held by the University of Texas at Austin. Please contact us if you would like to obtain a license to use RECIPE indexes in your commercial product.
We thank the National Science Foundation, VMware, Google, and Facebook for partially funding this project. We thank Intel and ETRI IITP/KEIT[2014-3-00035] for providing access to Optane DC Persistent Memory to perform our experiments.
Please contact us at [email protected] and [email protected] with any questions.