Automated Performance Control in a Virtual Distributed Storage

System

Report：Zhen Ming Wu

2008 IEEE 9th Grid Computing Conference

OutlineIntroduction

Virtual Distributed Storage System Architecture

Metadata

Storage machine

ISCSI server

Design Performance controller

Evaluation

Read and write performance

Performance control

Conclusion

Introduction

Strong demand for disk storage

Abundant yet idle resources on desktop

Machines

Various performance requirements for Storage

Service

Call Storage@desk

Virtual Distributed Storage System Architecture

Metadata

Correctly service clients’ requests.

Appears as a virtual volume

stores three types of metadata about the system

Storage machine

Only Storage machine

According to SAN Architecture

keeping the database updated

ISCSI server

act as the interface point

responsible for maintaining

enforcing access-control security policies

replication strategies ： RAID 1

Design Performance controllerBandwidth control with the help of a feedback controller

dynamically throttles

K： Time intervaly(k) ： sensor records the bandwidthr： reference bandwidth

e(k) = r – y(k)： control input

actuator will enforces a delay function : (u(k))

But a smaller error e(k+1) = r – y(k+1)

Proportional controller： e(k)control gain： Kpintegral controller： Kicontrol output at the time： u (k −1)

EvaluationStorage@desk

ISCSI server install Windows Server 2003 and MySQL database

Storage machine install Fedora 7 and kernel 2.6.23

Intel Xeon 8x CPUs at 2.33 GHz

16 GB RAM

250 GB SATA hard drive

Windows Server 2003

7 using kernel 2.6.23

Clients

Windows XP

P4 CPU at 2.4 GHz

512 MB RAM

100 Mb/s fast network

Create a NTFS file system on a virtual volume of 50 GB

Compare the performance of Storage@desk and Windows CIFS

Read and write performanceTest Read and write performance from 1MB to 2GB

Performance controlClient sequentially writes a 20 GB file

Client sequentially reads a 20 GB file

Two client reads and writes to random locations in a 2 GB file

Total runtime is 1,800 seconds

Divided into three regions of 600 seconds

Time interval of ten seconds record

writes reads

Conclusion

Utilizes distributed machines to create an extremely large storage pool

Storage@desk to provide predictable storage services

feedback control can automatically throttle

Thanks for your attention