Under construction!
Soon efficient support for Intel Optane will be added.
This is the experimental source release for Durable Masstree, a fast, multi-core key-value store that runs on NVM. Ensuring the recoverability of the data after crash is expensive due to the cost involved in ordering the writes to NVM. Durable Masstree offers a new design space, where the cost of ordering writes by using flushes and fences can be reduced through the use of fine-grain checkpointing and using a log that is stored within a cache line as described in Fine-Grain Checkpointing with In-Cache-Line Logging. Durable Masstree code offers the possibility to explore the described design space.
This document describes how to run Durable Masstree and obtain results.
- Jemalloc: It is preferable that you use jemalloc as allocator.
Get the sources using:
$ git clone https://github.com/epfl-vlsc/Masstree.git
Please make sure you have jemalloc installed. Afterwards, run the commands below to build the project.
$ ./bootstrap.sh
$ ./configure
$ make all
Load the kernel module for kernel flushes.
$ make flush_all
$ make flush_load
- In
kvthread_pallocator.hhplease changeDATA_BUF_SIZE, PDATA_FILENAME, PDATA_DIRNAMEand inincll_pextlog.hh, please change,PBUF_SIZE, PLOG_FILENAMEaccording to the mapping location and size of the memory available to your system.
Run the script for tests. Check incll_configs.hh for trying out different configurations.
$ chmod +x *.sh
Run a single ycsb experiment.
$ ./run_workload.sh ycsb_a_uni 10
Run many different experiments. Check the output folder for results.
$ ./run_this.sh
https://docs.google.com/spreadsheets/d/1pAZeWBC6P7nOY8_oDlYT_alZqMIIc0of9fWGUjBmW8U/edit?usp=sharing
Cohen, Aksun, Avni, Larus, "Fine-Grain Checkpointing with In-Cache-Line Logging," ASPLOS 2019.