PreemptDB a database engine that supports lightweight userspace preemptible transactions enabled by Intel x86 user interrupt (uintr). PreemptDB allows high-priority, short-transactions to quickly preempt and pause long-running, low-priority transactions which is resumed after the short-transaction concludes.
Check out more details in our SIGMOD 2025 paper (received Best Paper Award!). If you use our work, please cite:
Low-Latency Transaction Scheduling via Userspace Interrupts: Why Wait or Yield When You Can Preempt?
Kaisong Huang, Jiatang Zhou, Zhuoyue Zhao, Dong Xie, Tianzheng Wang.
SIGMOD 2025
Please build and run this tool to check if your CPU supports user interrupt.
We have two options to enable kernel support for user interrupts:
- Patched Kernel: We provide a patched kernel with uintr support. You can download it from here. Follow the instructions in the repository to build and install the kernel.
- Character Device: If you prefer not to use a patched kernel, you can use the character device interface. This allows you to enable uintr support without modifying the kernel. If your kernel already enables uintr, do not use the character device. We will release the character device driver soon.
PreemptDB uses huge pages for memory allocation. You can allocate huge pages using the following command:
sudo sysctl -w vm.nr_hugepages=[preferred number of huge pages]
sudo sysctl -pgit clone https://github.com/sfu-dis/preemptdb.git
cd preemptdb
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=[Debug|RelWithDebInfo|Release] # Default: Release
makeBefore you run PreemptDB, make sure you have chosen the desired parameter values:
--log_data_dir: Directory for log data. Make sure the directory exists and is writable.--node_memory_gb: Memory allocated for each node in GB. Make sure the total memory allocated does not exceed the available hugepage memory.--tpcc_scale_factor: Scale factor for the TPC-C benchmark. This determines the size of the database.--tpcc_workload_mix: [New-Order, Payment, Credit-Check, Delivery, Order-Status, Stock-Level, TPC-H Query 2 variant, unused profile]: TPC-C transaction profiles with additional analytical query profiles. It is the workload mix run in the default transaction context. We typically use 0,0,0,0,0,0,100,0 which means only long-running TPC-H Query 2 is run in the default transaction context.--tpcc_prioritized_workload_mix: TPC-C transaction profiles with additional analytical query profiles. It is the workload mix run in the prioritized transaction context. We typically use 50,50,0,0,0,0,0,0 which means 50% TPC-C New-Order and 50% TPC-C Payment are run in the prioritized transaction context.--scheduling_policy: 1: Vanilla (FIFO wait), 2: Preemptive (PreemptDB), 3: Cooperative (userspace transaction context switch but not coroutine-based).--prioritized_queue_size: Size of the prioritized/preemptive transaction queue in each worker thread.--global_queue_size: Size of the global preemptive transaction queue in the uintr sender (admission control) thread.--arrival_interval_us: Arrival interval of preemptive transactions in microseconds.--max_preempted_cycle_pct: Maximum percentage of cycles that can be preempted from the default transaction context.--latency_sample_size: Number of latency samples to collect for each run.
mkdir /home/$USER/preemptdb-log
cd build
./benchmarks/tpcc/tpcc_SI_sequential \
--log_data_dir=/home/$USER/preemptdb-log \
--null_log_device=1 \
--pcommit=0 \
--node_memory_gb=64 \
--tpcc_scale_factor=8 \
--tpcc_workload_mix=0,0,0,0,0,0,100,0 \
--tpcc_prioritized_workload_mix=50,50,0,0,0,0,0,0 \
--scheduling_policy=2 \
--prioritized_queue_size=8 \
--global_queue_size=64 \
--arrival_interval_us=1000 \
--max_preempted_cycle_pct=100 \
--latency_sample_size=1000 \
--seconds=30 \
--threads=8