Virtual Resource Management in PaaS cloud service

Today, I read intelligent management of receivalized resources for database systems in cloud environment from icde2011 and take some notes:

Outline: maximize the benefits of cloud service providers with comprehensive consideration of user satisfaction and system overhead. This article proposes a cost-based resource management system, smartsla. The system consists of two parts: the system model module and the resource allocation module. The former is mainly used to analyze various factors, such as the number of CPU shards, memory size, number of Database Replicas, query frequency, and other impacts on the virtual machines deployed by users (that is, the impact on user satisfaction ), in this paper, several methods of machine learning are used to obtain a function related to user satisfaction with various factors. The latter mainly considers how to rationally allocate resources among multiple users, especially a variety of different users (such as VIP users and ordinary users) to maximize the benefits of cloud service providers.

Detailed description:

In this article, the multi-user sharing level is private virtual machine, that is, each user's database is located on its own independent virtual machine, cloud service providers by adjusting the size of resources occupied by virtual machines to meet user needs. Multiple different virtual machines are located on one physical machine. Therefore, we need to reasonably allocate limited resources on the physical machine to meet the SLA of different users ). (In fig1, a user occupies multiple virtual machines, which are actually backups, that is, the MySQL slave is a backup of the MySQL master ).

A user's SLA can be expressed as a penalty segmentation function: when the response time (execution time) of a user's submitted task is less than a certain value, the cloud service provider does not need to pay a penalty, when the response time of a user's task exceeds a certain value, the cloud service provider will pay the user a certain penalty. The average value of the multi-user penalty segmentation function is called the average SLA penalty overhead (AC). The purpose of resource allocation is to minimize the amount of the average penalty.

This paper proposes a new system resource management system, smartsla, which consists of the following two modules.

1) System Model module: First, you need to find the relationship between the number of CPU shards, memory size, number of Database Replicas, and query frequency allocated to a user and the AC ). This function is obtained using machine learning. Linear regression, regression tree, and boosting approach methods were adopted respectively.

2) Resource allocation module: The system resource module focuses on the relationship between the AC and the number of CPU shards, memory size, database copy, and query frequency in a single user case. This module focuses on how to allocate resources among users of different levels to minimize the total amount of fines. You can assign different weights to the penalty functions of different levels of users and modify their critical response time. For example, the critical response time of a gold medal user can be relatively small and the amount of fines to be paid after the time is exceeded can be relatively large. In this way, various resources (such as the number of CPU shards, memory size, and number of Database Replicas) need to be allocated to different users, especially users of different levels, to minimize the penalty.

First, consider how to allocate the number of CPU shards and memory size. Obviously, when a user of a certain level queries a large frequency, the CPU and memory allocated by the user of this level should increase accordingly.

The next step is to consider the number of Database Replicas. intuitively, the higher the number of replicas, the shorter the access speed and latency, and the lower the penalty. However, increasing the number of replicas will increase the overhead of the system architecture (Infrastucture Cost ). This is obvious. To maintain more copies, more hardware and software equipment, management fees, and electricity fees are required. In this article, we assume that the overhead of each new copy is fixed. In addition to the system architecture overhead, when increasing the number of replicas, there will be a certain amount of Operation overhead (Action Cost), that is, the overhead for data migration, which also needs to be considered.

The experimental results show that using SmartSLA can effectively reduce the total cost of fines and maximize the interests of cloud service providers.

2012-1-13 Kemaswill