Eve: A MeasurementCentric Emulation Environment for Adaptive Internet Servers

1
Eve A Measurement-Centric Emulation Environment
for Adaptive Internet Servers
2
Needs of Service Emulation

• Performance evaluation
• How fast is my newest quad CPU server?
• Capacity planning
• How many servers do I need?
• Where is the bottleneck?
• Service adaptation evaluation
• When / where should services be migrated?
• What if static pages are used to deal with flash
  crowd?

3
To Simulate or Emulate?

• Simulate
• To represent certain features of the behavior of
  a system by the behavior of another system.
• Emulate
• To duplicate the functions of one system with a
  different system.
• Federal Standard 1037C Telecom Glossary

4
Why Not Using Benchmarks?

• Fixed workload
• Client model
• Server model
• Closed architecture
• Difficult to incorporate feedback control, etc.
• Lengthy setup
• Multiple servers (web, database, etc.)
• Specific data set

5
Goals of Emulation Environment

• Efficient
• Adaptive
• Extensible
• Accurate
• Scalable

6
Features of Eve

• Interactions between clients and servers
• Efficient data management
• Rapid development and deployment

7
Outline

• Eve Architecture
• DSV Distributed Shared Variable
• In-path Data Manipulation and Triggers
• FASE Feedback-triggered Adaptive Service
  Emulator
• Architecture
• Resource Demand Request
• Evaluation
• Conclusion

8
Architecture of Eve
Eve Module 1
Emulation Code
User-Level Threading Lib
DMM
Eve Module Substrate (EMS)
EveModule 2
EveModule N
DMM
DSV Proxy
EGS Proxy
DMM
Module Manager
EGS
DSV
DMM
DMM
Eve Kernel
9
DSV Distributed Shared Variables

• Location-independent
• For both configuration and measured data
• No enforced consistency model
• In-path data manipulation
• Pre-process data
• Post-process data

10
DMM Data Manipulation Modules

• Extend both Eve modules and kernel
• Create a shadow copy in Eve modules
• Override data read/write functions
• Triggered data synchronization
• Time
• Update count
• Threshold
• Reusability

11
DMM Example
Eve Module
Thread generatesdata
DSV intercepts data and redirects to DMM
DMM for var X
Shadow data
Wait forTrigger
Process dataand updateshadow copy
DSV Proxy
EMS
Sync withDSV in Eve kernel
12
FASE Feedback-triggered Adaptive Service
Emulator

• Client emulation
• Support of both RDR and HTTP requests
• Dynamic behavior adaptation
• Service emulation
• Workload described in RDR
• Virtual server scenario support
• Dynamic virtual-to-physical server mapping
• Service migration control

13
FASE Architecture
14
RDR Resource Demand Request
parallel fork
15
Eval DSV Response Time
DSV on local machineDSV on remote machineDSV
w/DMM on remote machine
Response Time (usec)
Read Write Create
GTAS GetInfo
DSV Operations
16
Eval Consistency in DMM
17
Eval Benefits of DMM
18
FASE Adaptation Scenario
Emulated server 1
Can be migrated to external CDN
Emulatedclients
Service 2 Web server w/ dynamic content
Can be replicated to another server
Emulated server 2
Replica of Service 2
19
FASE Adaptation Example
ClientServer 1Server 2
20
Conclusion

• Eve a scalable, extensible, and adaptive
  emulation environment
• FASE adaptive service emulation
• Extended to emulate SOA applications using BPEL
  transactions
• Questions?
• jamjoom_at_us.ibm.com chtsai_at_eecs.umich.edu