A Unified, Low-overhead Framework to Support Continuous Profiling and Optimization

1
A Unified, Low-overhead Framework to Support
Continuous Profiling and Optimization
22nd IEEE International Performance Computing and
Communications Conference (IPCCC2003)

• Xubin (Ben) He (hexb_at_tntech.edu)
• Storage Technology Architecture Research(STAR)
  Lab
• Department of Electrical and Computer Engineering

2
Outline

• Introduction
• Architecture and Design
• Performance Evaluations
• Conclusions and Future Work

3

• Introduction

4
Motivations

• System profiling is an important mechanism to
  observe system activities.
• Profiling-based optimization has become a key
  technique.
• Continuous and online optimization is needed
  because of changed system usage patterns.

5
Current State-of-the-art

• Traditional approaches bring high overhead to
  already overloaded system.
• Profiling and optimization overhead
• Raw Data Gathering
• Data Recording
• Data Processing
• Feedback

6
Introducing Unified, Low-overhead Framework (ULF)

• Offload computing overheads from host processors
  to an embedded processor
• Continuous feedback loop model
• 1. Low overhead profiling to gather system event
  data
• 2. Parallel processing raw data and policy
  generation
• 3. Apply policy to host

7

• Introduction
• Architecture and Design
• Performance Evaluations
• Conclusions and Future Work

8
Components

• ULF board an embedded processor with a
  sufficient amount of RAM
• Host-side module APIs as a library or kernel
  module
• Board-side moduleembedded os, a libray, plug-ins

9
ULF Board

• Low cost, low power embedded processor.
• Expandable with secondary PCI slot.
• Interface with host via standard PCI slot

10
Interaction between Plug-ins and Boards
Initial stage--gtRunning---gtCleanup
11
Example Applications

• Low overhead profiling
• On-line program optimizer
• On-line file system cache optimizer

12

• Introduction
• Architecture and Design
• Performance Evaluaitons
• Conclusions and Future Work

13
Experimental Setup

• Methodology
• Prototype using Intel IOP310 processor, Linux
  2.4.16
• I/O profiling tool LTT(Linux Trace Toolkit)
• Workloads
• Postmark of Network Appliances throughput
• 20k initial files, transactions ranging from 150k
  to 300k.
• Iozone
• 4 configurations
• NTNR Neither Traced Nor Recorded
• TNR Traced but Not Recorded
• TDR Traced and Disk Recorded
• TFR Traced and ULF Recorded

14
PostMark Results
15
Different W/R ratio
16
Iozone results
17

• Introduction
• Architecture and Design
• Performance Evaluations
• Conclusions and Future Work

18
Conclusions

• A unified, low-overhead framework helps profiling
  tools to save profiling data rapidly and perform
  run-time parallel processing.
• Reduces profiling overhead
• LTT 40--gt0.4.

19
Future Work

• Apply ULF to more case studies
• Performance
• Adaptively adjust system prefetching and caching
  policy
• Online code rewrite and recompilation
• Security
• Monitor abnormal system access and high risk
  events.
• Intrusion detection

20
Acknowledgements

• Dr. Ken Yang
• Ming Zhang
• NSF
• Manufacturing Center at T.T.U

21
A Unified, Low-overhead Framework to Support
Continuous Profiling and Optimization
IPCCC2003

• Xubin He (hexb_at_tntech.edu)
• http//www.ece.tntech.edu/hexb/starlab.htm
• Storage Technology Architecture Research(STAR)
  Lab
• Department of Electrical and Computer Engineering