Implemented on PostgreSQL 16.2 with the -DRCC compile flag. RCC on PostgreSQL supports Repeatable Read.
- Repeatable Read only. RCC on PostgreSQL tracks only commit dependencies. Serializable isolation (PostgreSQL's SSI) is not supported.
- No RCC-C. Columnar RCC needs redo logs with column-level update deltas, but PostgreSQL's WAL stores full tuple images. RCC on PostgreSQL therefore tracks dependencies at the record level.
- OS: Ubuntu 22.04 LTS
sudo apt-get update
sudo apt-get install -y make gcc libicu-dev pkg-config libreadline-dev zlib1g-dev numactl gnuplotBuild sysbench with PostgreSQL support:
sudo apt-get install -y make automake libtool pkg-config libaio-dev libpq-dev
cd /tmp
git clone https://github.com/akopytov/sysbench.git
cd sysbench
./autogen.sh
./configure --with-pgsql
make -j$(nproc)
sudo make installexport rcc_data_dir=/path/to/data # PostgreSQL data directory
export rcc_log_dir=/path/to/log # PostgreSQL log directoryNote: The experiment scripts erase both directories before every run. Use dedicated directories.
./run_server.sh [options]
Options:
--vanilla Build vanilla PostgreSQL (without RCC)
--rr Build RCC (-DRCC) [default]
--skip-build Skip build step (use existing binaries)
--numa <mode> Server NUMA policy: pin (node 0) [default], interleave, none
-h, --help Show helpExamples:
./run_server.sh # Build RCC and start
./run_server.sh --vanilla # Build vanilla and start
./run_server.sh --skip-build # Use existing build, just start server
./rcc_postgresql/bld/bin/pg_ctl stop -D $rcc_data_dir # Stop serverEach script sweeps three configurations over the client threads: it builds PostgreSQL, initializes and loads the database, warms up for 50 seconds (discarded) and measures 50 seconds.
bash scripts/ycsb_threads.sh # YCSB
bash scripts/tpcc_threads.sh # TPC-COptions:
--results-dir <path> Results directory (default: results/ycsb_threads, results/tpcc_threads)
Results are stored in results/:
results/tpcc_threads/
├── rcc_rr/
│ ├── s1_t1/
│ │ ├── summary.txt # configuration of the run
│ │ ├── sysbench.log # measured run
│ │ ├── warmup.log
│ │ ├── consistency.txt # TPC-C consistency check
│ │ └── postgresql.log
│ └── ...
├── vanilla_rc/
└── vanilla_rr/
results/ycsb_threads/ has the same layout with t<threads>/ run directories, a load.log per run and no consistency check.
grep "transactions:" results/tpcc_threads/rcc_rr/s1_t64/sysbench.log
grep "transactions:" results/ycsb_threads/rcc_rr/t64/sysbench.logThe plot scripts and their gnuplot templates are in plots/. Each one writes the throughput of every configuration (tps_128.csv) into the results directory and renders the graph there:
bash plots/plot_ycsb_threads.sh # results/ycsb_threads/ycsb_threads.pdf
bash plots/plot_tpcc_threads.sh # results/tpcc_threads/tpcc_threads.pdf| Experiment | Benchmark | Description |
|---|---|---|
| ycsb_threads | YCSB | 10M rows, zipf theta 0.9, 8 reads + 8 updates per transaction in ascending key order; threads 1, 16, 32, 64, 96, 128 |
| tpcc_threads | TPC-C | 1 warehouse, NewOrder:Payment = 1:1; threads 1, 16, 32, 64, 96, 128 |
| Mode | Compile Flags | Isolation |
|---|---|---|
| vanilla_rc | -O2 |
Read Committed |
| vanilla_rr | -O2 |
Repeatable Read |
| rcc_rr | -O2 -DRCC |
Repeatable Read |