Tabular: Efficiently Building Efficient Indexes

Tabular is a new parallel programming library that models data structures as ACID relational tables to provide concurrency and persistence transparently.

Tabular turns a single-threaded in-memory index into a concurrent, persistent counterpart without requiring the developer using parallel programming techniques, easing programming effort.

More details are described in our VLDB 2025 paper below (preprint here). If you use our work, please cite:

Tabular: Efficiently Building Efficient Indexes.
Ziyi Yan, Mohamed Farouk Drira, Tianxun Hu and Tianzheng Wang.
VLDB 2025

Setup

Hugepages

echo <number-of-1GB-hugepages> | sudo tee /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages

Dependencies

Build tools

• gcc
• cmake
• python
• cpplint

Libraries

• glog
• gtest
• tbb
• jemalloc

Build

mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release # RelWithDebInfo, Debug
make -j`nproc`

Index Benchmarks

Variants

Hand-crafted:

• btree_olc_bench_256
• hashtable_olc_bench_256
• art_olc_bench

Tabular:

• btree_inline_tabular_bench_256
• hashtable_inline_tabular_bench_256
• art_tabular_bench

Naive-OCC:

• btree_inline_tabular_materialized_bench_256
• hashtable_inline_tabular_materialized_bench_256
• art_tabular_materialized_bench

Naive-MVCC:

• btree_tabular_bench_256
• hashtable_mvcc_tabular_bench_256

Pessimistic Locking:

• btree_lc_stdrw_bench_256
• btree_lc_tbbrw_bench_256
• hashtable_lc_stdrw_bench_256
• hashtable_lc_tbbrw_bench_256
• art_lc_stdrw_bench
• art_lc_tbbrw_bench

Other indexes:

• bptree_bench_1K_16_16
• masstree_bench

Workload Types

• C: Lookup-only
• G: Lookup-mostly
• A: Balanced
• H: Update-mostly
• F: Update-only

Command-line interface

# Volatile
$BUILD_DIR/benchmarks/ycsb/ycsb_<variant> --workload <workload> --threads <number of threads> --records <number of keys> --duration <seconds>
# Persistent (only works for Tabular indexes)
$BUILD_DIR/benchmarks/ycsb/ycsb_<variant> --workload <workload> --threads <number of threads> --records <number of keys> --duration <seconds> --persistent --log-dir=<logging directory>

TPC-C Benchmarks

$BUILD_DIR/benchmarks/tpcc/tpcc_masstree_bench --duration <seconds> --threads <number of threads> --scale-factor <scale factor>
$BUILD_DIR/benchmarks/tpcc/tpcc_tabular_btree_bench --duration <seconds> --threads <number of threads> --scale-factor <scale factor>

Notes

To minimize NUMA effect, run benchmarks with numactl --membind 0 or numactl --interleave 0,1,..,<N> when threads are within one socket or across N sockets.

Every benchmark binary accepts --help to print complete available options.