SAC 2025 UnderReview ... 😵💫
We propose DATATune as a novel approach that differs from trditional database parameter tuning methods.
To minimize the time required for data generation, DATATune incorporates an augmentation method for small datasets. Moreover, it creates a latent space by reducing database parameter information and optimizing all parameters. Furthermore, by injecting external metric information in the latent space yield precise tuning results by directly incorporating information from the target workload.
Our results show that when comparing the performance improvements of DATATune and the baseline models across four different workloads on MySQL and RocksDB, we observed maximum performance improvements of 1332% for RocksDB and up to 11.82% in throughput and 46.01% in latency for MySQL.
This is our experiment workload information.
| Workload Index | Scale Factor | Data Size | Read | Insert | Scan | Update | Read Modify Write |
|---|---|---|---|---|---|---|---|
| A | 12000 | 15GB | 50% | - | - | - | 50% |
| B | 12000 | 15GB | 95% | - | - | 5% | - |
| E | 12000 | 15GB | - | - | 5% | 95% | - |
| F | 12000 | 15GB | 50% | - | - | - | 50% |
| Workload Index | Value Size | # of Entry | READ | WRITE | UPDATE |
|---|---|---|---|---|---|
| R90W10 | 16384 | 65472 | 90% | 10% | X |
| R50W50 | 16384 | 65472 | 50% | 50% | - |
| R10W90 | 16384 | 65472 | 10% | 90% | - |
| UPDATE | 16384 | 65472 | - | - | O |
(The 'sac experiment' file is the code that performed the ablation study.)