PrediCache

This repository contains the PrediCache prototype for the paper "Predictive Translation: High-Performance Buffer Management Without the Trade-Offs" (accepted at SIGMOD 2026).

Building and Running the Prototype

The prototype can be built using make and run by executing the built predicache binary.

Configuration

The prototype can be configured using the following environment variables:

• BLOCK: storage block device (e.g. /dev/nvme0n1 or /dev/md0); default=/tmp/bm
• POOLGB: size of the buffer pool in GB; default=1
• BATCH: batch size for eviction in pages; default=64
• RUNFOR: benchmark run duration in seconds; default=30
• WORKLOAD: one of 'tpc-c', 'random-read', 'ycsb'; default='tpc-c'
• THREADS: number of threads; default=1
• DATASIZE: number of warehouses for TPC-C, number of tuples for random read/ycsb benchmark; default=10

YCSB Specific Configuration

The payload size is fixed at 120 bytes and can only be changed by modifying the source code.

• ZIPF_FACTOR: zipf factor for YCSB workload, default=0.9
• OPS_PER_TX: number of operations per transaction for YCSB workload, default=1
• READ_RATIO: read ratio for YCSB workload, default=100 (100% reads)

Example Usage

To run the prototype with a TPC-C workload with 100 warehouses using 4 threads and a 2GB buffer pool, you can use the following command:

make
BLOCK=/dev/nvme0n1 POOLGB=2 THREADS=4 WORKLOAD=tpc-c DATASIZE=100 ./predicache

To run the prototype with a random read workload with 1000000 entries on a 128GB buffer pool using 64 threads, you can use the following command:

make
BLOCK=/dev/nvme0n1 POOLGB=128 THREADS=64 WORKLOAD=random-read DATASIZE=1000000 ./predicache

Dependencies

The prototype requires the following dependencies:

• gcc for compilation (clang might work but wasn't tested)
• make for building the project
• libaio-dev for asynchronous I/O operations
• liburing-dev for IO-uing support (optional, build with make uring)