• Summary
• Questions
• Introduction
  • Background & Motivation
  • Goal
  • Challenge
• Details
  • SLA definition:
  • Page replacement

Summary

Why need a shared buffer pool to reduce the cost? Resource solution and shared resource in a single DB server?

Questions

Introduction

Background & Motivation

Database as a service combines RDBMS with the goal of the reduced total cost.

Multitenance is crucial for service providers to increase consolidation and reduce costs.

• Multiple tenant databases are hosted inside a single database server, thus sharing the same DB server resources. As a result, one user’s workload may influence another user.
• Reserved resources to a tenant, isolating one tenante’s performance from the resource demands of another.

Goal

This paper studies the problem of how to share buffer pool memory in a multi-tenant setting effectively.

Challenge

1. Define accountability of service providers. The paper defines the metrics HRD: the difference between the hit ratio of baseline (memory is statically reserved to the user) and of shared-buffer-pool env. If the service provider is unable to meet this promise, a penalty function
2. How to effectively measure the HRD with low memory and CPU overheads. (baseline hit ratio requires some logic)
3. Design the page replacement algorithm. MT-LRU

Details

SLA definition:

• Hit ratio degradation will reduce the user’s experience. Each user will have
• Penalty function: how a service provider is accountable for any value of HRD

Page replacement

LRU: some pages were visited once, and then the previously least-recent-used page was dropped.

• But which would then be referenced again, resulting in increased I/O.
• LRU is unable to differentiate between pages that have relatively frequent references.
• In general, bad for database applications

LRU- K:

• visiting a page -> putting page to the history list
• Once Key’s visiting time in the history list exceeds K, -> putting Key to cache.
• If the cache is full, remove the Key whose kth timestamp has the largest distance with now.
• history also cannot grow infinitely. It uses FIFO、LRU、LFU to control.

MT-LRU:

• Alg objective: This paper designs an algorithm that would reduce to LRU-K for any individual tenant while being able to handle new requirements arising from multiple tenants with SLAs, including:
  • Asymmetry of tenants. Tenants could differ in the amount of memory promised in the SLA and their penalty functions. Furthermore, different tenants could be active at other times.
  • Unlike LRU-K, which aims to maximize the likelihood of a page being found in the cache, we need to optimize for different objectives, e.g., minimizing total penalties..=> Convert to minimize the number of misses.
• LRU-K cannot meet the requirements
• Algorithms.
  • Mark each page’s importance with the proposed pental function’s score instead of K’ visited timestamp.
  • Each time eviction will reduce the score, such that the older one has a low score -> consistent with LRU.
  • Eviction:
    • For each tenant, assign a score for each page.
    • Sort global pages, and get the percentile.
    • For each tenant, remove the pages with scores lower than the percentile.
    • update the score vector.

---

• ← Previous Post
• Next Post →