This repository exposes part of the test scripts, related codes and public workload traces for running core tests on the StRAID prototype.

Hardware Dependencies

We recommend deploying at least four dedicated NVMe SSDs for building StRAID in the experimental platform. These SSDs should to be installed on the PCIe slots belonging to a same CPU socket. Please note that the use of hardware and the installation method might produce different experimental results.

Software Dependencies

Run the script to install dependencies

cd code
./install_depends.sh 

or install dependencies as follows

sudo apt update
sudo apt upgrade
sudo apt install cmake
sudo apt install build-essential
sudo apt install autoconf
sudo apt install yasm
sudo apt install nasm
sudo apt install libtool-bin
sudo apt install libtbb-dev
sudo apt install libcds-dev
sudo apt install dstat

Install submodules

In the sections of motivation and evaluation in our paper, we used some tools for analysis and comparison, so we introduced them as submodules. To execute these parts of the code, please install the submodules.

git submodule update --init --recursive

Intel isa-l

https://github.com/intel/isa-l

 cd include/isa-l
 ./autogen.sh
 ./configure
 make
 sudo make install

IO_uring

https://github.com/axboe/liburing

 cd include/liburing
 ./configure
 make
 make install

libcuckoo

https://github.com/efficient/libcuckoo

cd include/libcuckoo
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=/usr/local -DBUILD_EXAMPLES=1 -DBUILD_TESTS=1 ..
make all
make install

Evaluation

We made scripts for two main test cases, including basic write/read tests and a trace-driven test. The first tests generate fixed numbers of requests with various I/O sizes and access patterns on StRAID. The trace-driven test will replay block-level I/O traces recorded in the trace file.

Please note that StRAID is currently a prototype and may crash unexpectedly. If you experience a crash or get stuck, please repeat the experimental program or script several times. If the test program keeps failing, please check the system configuration and input parameters are correct.

Micro-benchmarks

1. Set environment

   The environment variable like the number of instances (NUM_THREADS) and the size of dataset (DATASET_SIZE) should be set at the compile time, detailed are shown as follows:

   vim ./include/define.h
   
   // The number of running I/O-issuing Threads (IOT) and Workering Threads (WT)
   #define NUM_THREADS (16) 
   
   // The dataset size and the number of repeating workload executions
   #define DATASET_SIZE (1 * GB)
   #define LOOP (1)    
   
   // The I/O size of each read or write requests
   #define IO_SIZE PARTSIZE // IO size for partial-stripe write
   #define IO_SIZE FULLSIZE // IO size for full-stripe write
   
   // Number of data chunks and parity chunks in a stripe, which have to be consist with the number of deployed SSD devices. For example, a 5+1 RAID-5 should set the DATACHUNK_NUM to 5 and the PARITYCHUNK_NUM to 1
   #define DATACHUNK_NUM (5)                              
   #define PARITYCHUNK_NUM (1)                    
   

   Besides, the path of SSD devices and/or MD RAID should be assigned to the program.

   vim ./include/main.cc
   vim ./include/bench.cc
   
   // Assign paths of SSD devices used in experiment
   string lfile0 = "/dev/nvme0n1";
   string lfile1 = "/dev/nvme1n1";
   string lfile2 = "/dev/nvme2n1";
   string lfile3 = "/dev/nvme3n1";
   string lfile4 = "/dev/nvme4n1";
   string lfile5 = "/dev/nvme5n1";
   
   

2. Compile StRAID prototype test program in the default settings.

   cd code
   make -j 4
   

3. Config and run micro-benchmark test script

   This script first compiles StRAID based on the benchmark conditions assigned by user, and then launches test with four types of workload patterns (i.e., sequential/random Read and sequential/random Write).

   vim ./run_bench.sh
   
   # set this array to modify the number of launched I/O threads
   thread=(1 2 4 8 16 32 64) 
   
   # set this variable to modify the chunk size in StRAID to (chunksize * KB)
   chunksize=64 
   
   # set the variable to test full-stripe (FULLSIZE) or partial-stripe (PARTSIZE) write
   IOsize="FULLSIZE" 
   

   ./run_bench.sh
   

   Use the dstat tool to monitor disk I/O

   dstat -d -D "directories of SSD devices"
   # dstat -d -D /dev/nvme0n1,/dev/nvme1n1
   

4. The throughput and latency results are shown in:

   ./results/bench_out.txt
   

   For each benchmark pattern, the program will output its throughput and latency performance separately, including:

   • Average Throughput
   • IOPS
   • Average Latency
   • 50th-percentile Latency
   • 90th-percentile Latency
   • 99th-percentile Latency
   • 999th-percentile Latency

Macro-benchmarks

This repository exposes one typical block traces from Filebench for testing Linux MD and StRAID.

1. Set environment

   vim ./run_tracemd.sh
   
   thread=8          # set the number of user threads
   mdtype="5"        # set RAID level
   mdpath="/dev/md5" # set RAID path
   

   vim ./run_tracest.sh
   
   thread=8    # set the number of user threads
   chunksize=8 # set the chunk size of StRAID to (chunksize * KB)
   

2. Compile test program

   make -j 4
   

3. Launch marco-benchmark

   # marco-benchmark on Linux MD
   ./run_tracemd.sh
   
   # marco-benchmark on StRAID
   ./run_tracest.sh
   

4. The throughput and latency results are shown in:

   # results of Linux MD
   ./results/MD_trace_results.txt
   
   # results of StRAID
   ./results/ST_trace_results.txt
   

   The results contain:

   • Total throughput per second
   • Write throughput per second
   • IO per second (IOPS)
   • Latency CDF