Sheriff: A Regional Pre-Alert Management Scheme in Data Center - - PowerPoint PPT Presentation

Sheriff: A Regional Pre-Alert Management Scheme in Data Center Networks

Xiaofeng Gao, Wen Xu, Fan Wu, Guihai Chen and Ding-Zhu Du Shanghai Jiao Tong University

Background Data Center Networks

Background

• Goal of a data center network management

system is to be:

– Stable – Effective – Robust

• Several problems of existing management

schemes:

– Weakness of a centralized controller – Short-sighted mechanism

Introduction Centralization vs Distribution

• Drawbacks of Centralization

– Sharply increased response time – Upgraded or heterogeneous components

• What we need

– Distributed managers – Regional self-automatic control

Introduction Contingency vs Pre-Control

• Drawbacks of Contingency

– Working after detecting errors – Harmful to device prevention and system maintenance

• What we need

– Take early warnings – React in advance to avoid congestions

Introduction

• Sheriff

– Distributed (at end host side) – Pre-alert – Regional self-automatic

• Two phases

– PREDICTION

• ALERT message

– MANAGEMENT

• VM migration

System Design

Two kinds of graphs in a DCN

• 1. Wired Network Graph
• 2. Dependency Graph

Problem Formulation Problems & Solutions

• Overloaded servers

– Migrate VMs – Reshuffle VMs

• Congestion

– Check QCN – Modify the rate at end host

Problem Formulation Pre-alert & Actions

• Monitored servers by shim

– Information collection – Prediction by ARIMA model and NN model – Report ALERT value once it exceeds THRESHOLD

• Monitored switches by shim

– Flow congestion detection – Signal congestion flows

Problem Formulation Pre-alert & Actions

• Alert from servers or from ToR switches

– VM migration

• Alert from outer switches

– Flow reroute

• Implement flow reroute first
• VM migration

– More expensive – Slower

Problem Formulation Cost of VM Migration

• Six steps

– Initialization – Reservation – Iterative copy – Stop&Copy – Commitment – Activation

Problem Formulation Cost of VM Migration

• Cost of initialization
• Cost of transmission

– – Transmission time: – Utilization rate of the bandwidth:

• Cost of dependency

– – Unit cost per distance: – Physical distance of e:

Problem Formulation Cost of VM Migration

• Total cost:
• Gz:
• Goal: minimize

Problem Formulation Pre-alert Mechanism

• Collecting Necessary Information

– Workload profile – Normalized to [0, 1]

Problem Formulation Pre-alert Mechanism

• Time series prediction

– Autoregressive Integrated Moving Average (ARIMA)

• Modeling linear, dynamic signals

– Nonlinear Autoregressive Neural Network (NARNET)

• Modeling nonlinear, dynamic and chaotic
• Dynamic Model Selection

– For each method f – Choose method f which has the minimum value

Problem Formulation Alert Scheme

• Seriousness of the condition
• Collect ALERTs

– VM Migration

Alert-Migration Algorithm

Simplification of VM Migration Algorithm

• Cost of migration
• Simplification

– First step – Second step

• All pair shortest path problem
• K-median problem

– Only related with the source and destination of the migration

Alert-Migration Algorithm Framework

• Running periodically T time
• Each round

– Collect alerts – Select a group of candidate VMs (as sources) – VM Migration

• Not all VMs are migrated

– Parameters as portion for migration

Alert-Migration Algorithm Select Subroutine

• Remove delay-sensitive flows
• Pick up as many VMs with lowest value

– Dynamic Knapsack Algorithm

• If the priority parameter is one

– Pick VM with highest ALERT value

Alert-Migration Algorithm Migration

• Find optimal pairs: K-center problem

– Local Search Algorithm – It has an approximation ratio 3 + 2 / p with time complexity O(n^p)

• ACKs from the destination’s delegation node

– Enough capacity

• REJECTs from the destination’s delegation

node

– Recalculate possible migration destinations

Evaluation Network Trace Training

• Data from ZopleCloud Corp.
• Combined model has a smaller minimum

square error.

Evaluation Simulation For VM Migration

• Fat-tree & Bcube
• Workload balancing

Evaluation Simulation For VM Migration

• Result & Time Complexity

Conclusion

• Sheriff: A fast distributed pre-alert manage-

ment scheme in data center network

– Monitor locally – Predict possible ALERTs – Apply Flow Reroute / VM migration

• Evaluation

– Accuracy of the prediction – Efficiency of the migration algorithm

• Thanks for your attention!