DBMS Machine Provisioning

  • ICAC'11 - A Bayesian Approach to Online Performance Modeling for Database Appliances using Gaussian Models
    • Problem: modeling DBMS workloads using Gaussian Process
    • Keys:
      • It only focuses on modeling workloads and does not propose any application.
      • It demonstrate GP can predict quite accurate with small data set.
  • SIGMOD'18 - P-Store: An Elastic Database System with Predictive Provisioning
    • Motivation: previous work always react after an overloaded event happens.
    • Problem: they propose to make machine provisioning decision by predicting the following workloads.
    • Keys:
      • The target workload must be easy to predict.
      • There can be a few distributed transactions.
  • IEEE CloudCom'18 - DERP: A Deep Reinforcement Learning Cloud System for Elastic Resource Provisioning
    • https://ieeexplore.ieee.org/abstract/document/8590989
    • Motivation: previous work can not deal with large input space so we need a learning based method.
    • Method: Uses a DQN RL agent to decide when to add/remove machines and how many machines are added/removed to a DBMS cluster.
    • Key Points:
      • Targeting on NoSQL systems.
  • VLDB'19 - iBTune: individualized buffer tuning for large-scale cloud databases
  • VLDB'21 - Seagull: An Infrastructure for Load Prediction andOptimized Resource Allocation
    • http://www.vldb.org/pvldb/vol14/p154-poppe.pdf
    • Problem: to predict the load of a DBMS server and use the predicted info to decide when to backup the DB.
    • Keys:
      • Focuses on system design
      • Assumes the target workload have periodical patterns
      • Tried methods to predict workloads
        • Singular Spectrum Analysis
        • Feed-forward Networks
        • Prophet: a software with a model proposed by Facebook to predict time series data with yearly, weekly, and daily patterns.