RCC is a Redo log-based Concurrency Control scheme for high write contention. This repository implements it on MySQL, PostgreSQL and DBx1000.

With in-place updates, write-write conflicts serialize transactions that update the same record. RCC instead:

• updates a record out of place, as a speculative version built from the redo log entry, so conflicting transactions do not wait for their predecessors to commit;
• guarantees serializability with lock-based conflict ordering and a dependency graph;
• installs updates at commit, so an abort only discards speculative versions.

On top of this, RCC introduces two techniques:

• Optimistic deadlock handling grants locks without deadlock checks and detects cycles asynchronously.
• Columnar RCC (RCC-C) removes record-level false conflicts with column-level redo logs.

Repository layout

• MySQL/ — MySQL with RCC (see MySQL/README.md)
• DBx1000/ — DBx1000 with RCC (see DBx1000/README.md)
• PostgreSQL/ — PostgreSQL 16.2 with RCC (see PostgreSQL/README.md)
• plots/ — plot scripts and gnuplot templates of the paper's figures

Each system's README covers its setup, build, experiments and results.

Correctness: All TPC-C runs on MySQL and PostgreSQL passed the BenchmarkSQL consistency checks (scripts/consistency_check.sql). DBx1000 runs the same checks when built with TPCC_CONSISTENCY true (details).

Testing Configuration (from paper experiments):

• AMD EPYC 9754 CPU (128 cores at 2.25GHz)
• 256GB DRAM
• Ubuntu 22.04

Important: Set the CPU to maximum performance before running any experiment. By default the schedutil governor lowers the clock under light load, and idle cores enter deep C-states (C2 on the EPYC 9754: 800 µs exit latency). Both slow down lock handoffs and make throughput vary between runs. These settings reset on reboot, so apply them again after every boot.

cd /sys/devices/system/cpu
echo performance | sudo tee cpu[0-9]*/cpufreq/scaling_governor > /dev/null
echo 1 | sudo tee cpu[0-9]*/cpuidle/state[2-9]/disable > /dev/null

cat cpu*/cpufreq/scaling_governor | sort | uniq -c
grep . cpu0/cpuidle/state*/name cpu0/cpuidle/state*/disable

PostgreSQL has separate YCSB and TPC-C experiments; see PostgreSQL/README.md.

MySQL Fig. 4–6 use SERIALIZABLE isolation, 30s warmup and 30s measurement. After building the engine/client binaries, create reusable data and run:

cd MySQL
bash scripts/load_sysbench_data.sh
bash scripts/load_tpcc_data.sh
bash scripts/fig4_sysbench_threads.sh
bash scripts/fig5_tpcc_threads.sh
bash scripts/fig6_sysbench_rows.sh

Each figure supports --dry-run to inspect its plan. See MySQL setup and options.

plots/ has a plot script per figure (plot_<fig>.sh) and the gnuplot templates (templates/). Each script reads the results directory of its experiment and writes the plotted data and the PDF there. It needs gnuplot (sudo apt-get install -y gnuplot).

bash plots/plot_fig7a.sh               # DBx1000/results/Fig7a/fig7a_ycsb.pdf
bash plots/plot_fig7b.sh               # DBx1000/results/Fig7b/fig7b_tpcc.pdf
bash plots/plot_fig8.sh                # DBx1000/results/Fig8/fig8a_tps.pdf, fig8b_time.pdf
bash plots/plot_fig7a.sh /path/to/results/Fig7a   # another results directory