Efficient Compactions Between Storage Tiers with PrismDB (ASPLOS 2023)

This repository contains source code and instructions to reproduce key results in Efficient Compactions between Storage Tiers with PrismDB paper from ASPLOS 2023.

• Contacts: Ashwini Raina ([email protected]) and Jianan Lu ([email protected])

Getting the Source

git clone --recurse-submodules https://github.com/princeton-sns/prismdb.git

Dependencies

• sudo apt update
• sudo apt install python3-pip
• pip3 install -U PyYAML (needs YAML version 5.1 or later)
• sudo apt-get install cgroup-tools
• sudo apt-get install cmake
• sudo apt-get install libsnappy-dev
• sudo apt install libtbb-dev
• sudo apt-get install google-perftools (note: if libtcmalloc is not installed in /usr/lib then run this command (with the correct libtcmalloc version): sudo ln -s /usr/lib/x86_64-linux-gnu/libtcmalloc.so.<insert version here> /usr/lib/libtcmalloc.so)

Experiment Configuration

Setting paths for optane and flash disk

• Create a folder prism on the optane ssd.
• Create a folder prism on the flash ssd.

Configure the experiment parameters in scripts/config_test_example.yml file

• mem_alloc: cgroup dram allocation
• cpu_alloc: cgroup cpu allocation
• ssd: "het" for optane+flash setting, "optane" for optane only and "qlc" for flash only
• optane_ratio: optane to flash ratio
• pop_cache_size: clock cache size in terms of keys
• pop_cache_thresh: clock cache optane pinning threshold
• num_partitions: number of partitions
• db_size: database size in terms of keys
• key_size: key size in bytes
• value_size: value size in bytes
• num_clients: number of database clients (NOTE: it should be same as number of partitions)
• workload: ycsb workload "ycsba", "ycsbb", "ycsbc", ycsbd", "ycsbe", "ycsbf"
• distribution_param: (reads distibution, writes distribution). 0 stands for uniform and -1 stands for both read and write coming from the same distribution. distribution_param [(0.99 -1)] means both reads and writes follow Zipfian 0.99
• read_ratio: reads to writes ratio
• workload_operations: number of operations
• warmup_ratio: warmup ratio

Building and Running

Compile and run the experiment

• On terminal, go to the scripts folder and run sudo python3 run.py <experiment config file> <log folder name>
• e.g. sudo python3 run.py config_test_example.yml ycsb_log
• Logs for this experiment will get stored in scripts/logs/ycsb_log folder
• Results are stored in scripts/logs/ycsb_log/output.log. Throughput and latency results can be extracted from the output.log

Throughput:

• From the scripts folder run ./report.sh <output log path> <num_partitions>
• e.g. ./report.sh scripts/logs/ycsb_log/<experiment_folder>/output.log 8 will print total throughput (kops/s)

Latency:

• grep for "LATENCY HISTOGRAM" in output.log. It reports two instances - one for the loading phase and one for the running phase.

Parameter Tuning for Promotions

Read heavy workloads can benefit from PrismDB promotions. Promotions need some parameter tuning in file db/db_impl.h.

• read_dominated_threshold - for what read/write ratio should promotions be enabled? (default value is 0.95).

Promotions need some time to take effect. Evaluation of read-heavy workloads (like ycsb B and C) should be done with a longer workload (> 200 million operations).

Limitations

• Delete(key) api is buggy

Paper

• https://dl.acm.org/doi/10.1145/3582016.3582052