DistributedSystemResearch/23SIGMOD-polaris

Source code for the SIGMOD '23 paper “Polaris: Enabling Transaction Priority in Optimistic Concurrency Control”

DBx1000-Polaris

Polaris is an optimistic concurrency control algorithm with priority support. This repository implements Polaris on top of DBx1000 and DBx1000-Bamboo.

• DBx1000: Staring into the Abyss: An Evaluation of Concurrency Control with One Thousand Cores. Xiangyao Yu, George Bezerra, Andrew Pavlo, Srinivas Devadas, Michael Stonebraker.
• Bamboo: Releasing Locks As Early As You Can: Reducing Contention of Hotspots by Violating Two-Phase Locking. Zhihan Guo, Kan Wu, Cong Yan, Xiangyao Yu.

These repositories implement other concurrency control algorithms (e.g., No-Wait, Wait-Die, Wound-Wait, Silo) as the baseline for Polaris evaluation.

Quick Start: Build & Test

To test the database

python3 test.py experiments/default.json

The command above will compile the code with the configuration specified in experiments/default.json and run experiments. test.py will read the configuration and the existing config-std.h to generate a new config.h.

You can find other configuration files (*.json) under experiments/.

Advanced: Configure & Run

The parameters are set by config-std.h and the configuration file. You could overwrite parameters by specifying them from the command-line.

python3 test.py experiments/default.json COMPILE_ONLY=true

This command would only compile the code but not run the experiment.

Below are parameters that affect test.py behavior:

• UNSET_NUMA: If set false, it will interleavingly allocate data. Default is false.
• COMPILE_ONLY: Only compile the code but not run the experiments. Default is false.
• NDEBUG: Disable all assert. Default is true.

Below is a list of basic build parameters. They typically turn certain features on or off for evaluation purposes. The list is not exhaustive and you can find more on config-std.h.

• CC_ALG: Which concurrency control algorithm to use. Default is SILO_PRIO, which is an alias name of Polaris*.
• THREAD_CNT: How many threads to use.
• WORKLOAD: Which workload to run. Either YCSB or TPCC.
• ZIPF_THETA: What is the Zipfian theta value in YCSB workload. Only useful when WORKLOAD=YCSB.
• NUM_WH: How many warehouses in TPC-C workload. Only useful when WORKLOAD=TPCC.
• DUMP_LATENCY: Whether dump the latency of all transactions to a file. Useful for latency distribution plotting.
• DUMP_LATENCY_FILENAME: If DUMP_LATENCY=true, what's the filename of the dump.

*Fun fact: Polaris is implemented based on Silo but with priority support, so it was previously termed SILO_PRIO. The name POLARIS came from a letter rearrangement of SILO_PRIO with an additional A.

Below is another list of build parameters introduced for Polaris:

• SILO_PRIO_NO_RESERVE_LOWEST_PRIO: Whether turn on the lowest-priority optimization for Polaris. Default is true and it should be set true all the time (unless you want to benchmark how much gain from this optimization).
• SILO_PRIO_FIXED_PRIO: Whether fix the priority of each transaction. If false, Polaris will assign priority based on its own policy.
• SILO_PRIO_ABORT_THRESHOLD_BEFORE_INC_PRIO: Do not increment the priority until the transaction's abort counter reaches this threshold.
• SILO_PRIO_INC_PRIO_AFTER_NUM_ABORT: After reaching the threshold, increment the priority by one for every specified number of aborts.
• HIGH_PRIO_RATIO: What's the ratio of transactions that start with high (i.e., nonzero) priority. Useful to simulate the case of user-specified priority.

There are other handy tools included in this repository. experiments/*.sh are scripts to reproduce the experiments described in our paper. parse.py will process the experiment results into CSV files and plot.py can visualize them.