Governor: Autonomic Throttling for Aggressive Idle Resource Scavenging

1
Governor Autonomic Throttling for Aggressive
Idle Resource Scavenging

• Jonathan Strickland (1)
• Vincent Freeh (1)
• Xiaosong Ma (1 2)
• Sudharshan Vazhkudai (2)
• (1) Department of Computer Science, NC State
  Univ.
• (2) Mathematics and Computer Science Division,
  Oak Ridge National Laboratory

2
Presentation Roadmap

• Introduction
• Model and approach
• System implementation
• Performance results
• Conclusion and future work

3
Aggregating Desktop Computer Resources

• Personal computers pervasive
• Easily updated and well equipped
• Under-utilized
• Consolidate scattered resources by resource
  scavenging (resource stealing)
• Computing resources
• Condor, Entropia
• SETI_at_home, Folding_at_home
• Creating massive compute power
• Storage resources
• Farsite, Kosha, FreeLoader
• Aggregate distributed spaces into
• shared storage

(Courtesy SETI_at_home)
(Courtesy Folding_at_home)
4
Impact on Workstation Owners

• Foremost concern of resource donors
• Security and privacy impact
• Virtual machine/sandbox solutions
• Performance impact
• Existing approaches often too conservative
• Stop approach
• Stop scavenging when user activity detected
• Unable to utilize small pieces of idle time
• Does not overlap scavenging with native workload
• Priority-based approach
• Works for cycle-stealing
• Implicit, best-effort
• Range and granularity limited by operating system

5
Objectives and Contributions

• Goal systematic performance impact control
  framework
• Contributions Governor
• Explicit, quantified approach toward performance
  impact control
• Extensible framework for arbitrary scavenging
  applications and native workloads
• User-level, OS-independent implementation
• Evaluation with two types of scavenging
  applications

6
Presentation Roadmap

• Introduction
• Model and approach
• System implementation
• Performance results
• Conclusion and future work

7
System Entities

• Active on donated workstations
• Resource scavenging application (scavenger)
• Native workload
• Governor process
• Controls execution of scavenger
• Limits impact on native workload to target level
  a (e.g., 20)

8
Performance Impact

• Performance impact
• Caused by resource scavenging application on
  workstation owners native workload
• Metrics slow-down factor
• (Timescavenged Timeoriginal) / Timeoriginal
• May not reflect resource owner perceived impact
• Main approach resource throttling
• Throttle level (ß, 0ltßlt1)
• Timescavenging / Timetotal
• Major challenge to select appropriate ß value

9
Impact Benchmarking

• Characterize scavenger S against system resources
• Native workload as combination of resource
  consumption components
• Resource vector
• R (r1, r2, , rn)
• Benchmark vector
• B (B1, B2, , Bn)
• Measure S impact on Bi with various throttle
  levels
• Store impact curve
• Calculate target throttle level ßi with given
  impact level a

10
Native Workload Monitoring

• Native workloads typically complex and dynamic
• Online workload monitoring
• Activate corresponding ß when non-trivial native
  resource consumption detected
• Resource trigger vector
• ? (t1, t2, , tn)
• For each resource Ri
• ßi
• Overall ß min (ß1 , ß2 , ßn )
• Picking most restrictive ß across resources

ßi, if consumption ti
1, if consumption lt ti
11
Governor Architecture
0. impact benchmarking
User
target ?
system resources
Resource vectors (b1, b2 , ...) (t1, t2 , ...)
1. monitor resource activity
scavenger
3. throttle scavenger
2. compute overall b

• Adaptive
• Extensible and generic

Governor
12
Presentation Roadmap

• Introduction
• Model and approach
• System implementation
• Performance results
• Conclusion and future work

13
Dynamic Throttling Mechanism

• Fixed throttle interval I
• 1 second in our implementation
• Within each I, Governor
• Runs scavenger application for ßI
• Monitors native workload during (1-ß)I
• Adjust ß for next I

14
Resource Usage Monitoring and Triggers

• At beginning and end of each monitoring phase
  (1-ß)I
• Monitor resource usage
• CPU /proc/stat (cycles)
• Disk /proc/partitions (blocks)
• Network /proc/net/dev (bytes)
• Triggers (t array)

Resource Trigger value (t)
t CPU 1 utilization
t IO 0
t network 0
15
Presentation Roadmap

• Introduction
• Model and approach
• System implementation
• Performance results
• Conclusion and future work

16
Applications, Benchmarks, and Configurations

• Scavenger applications
• SETI_at_home
• Search for signals in slices of radio telescope
  data
• Computation-intensive
• FreeLoader
• Prototype for aggregating storage in LAN
  environments
• I/O- and network-intensive
• Single-resource benchmarks
• CPU EP from NAS benchmark suite
• I/O large sequential file read
• Network repeated downloading with wget
• Linux workstation
• 2.8GHz Pentium 4, 512MB memory, 80GB disk

17
Impact Benchmarking Results
SETI
FreeLoader
Resource Impact level (a) Impact level (a) Impact level (a) Impact level (a)
Resource 0.05 0.10 0.20 0.25
ßCPU 0.02 0.05 0.10 0.2
ßIO 1.0 1.0 1.0 1.0
ßnetwork 1.0 1.0 1.0 1.0
Resource Impact level (a) Impact level (a) Impact level (a) Impact level (a)
Resource 0.05 0.10 0.20 0.25
ßCPU 0.30 0.40 0.70 0.90
ßIO 0.05 0.10 0.20 0.25
ßnetwork 0.10 0.20 0.30 0.50
18
Multi-resource Workload Kernel Compile
Impact on native workload
Impact on scavenger app.
19
Synthetic Composite Workload

• Simulate common intermittent user activities
• Short sleep time between operations
• Writing 80MB data to file
• Browsing arbitrary directories in search of file
• Compressing data written previously and send via
  networks
• Browsing more directories
• Removing files written
• Takes about 150 seconds without concurrent user
  load

20
Composite Exec. Time and Impact

• Combine impact benchmarking results with
  real-time monitoring of composite workload
• Governor closely approximates target performance
  impact (a)

Resource Impact level (a) Impact level (a) Impact level (a) Impact level (a) Impact level (a) Impact level (a)
Resource 0.0 0.05 0.10 0.20 0.25 1.0
SETI_at_home impact 142 0 148 4.0 154 8.4 168 18.5 180 26.8 261 83.8
FreeLoader impact 142 0 150 5.6 157 10.6 172 21.1 180 26.8 211 48.6
21
Comparison with Priority Based Method (SETI_at_home)
22
Comparison with Priority Based Method (FreeLoader)
23
Presentation Roadmap

• Introduction
• Model and approach
• System implementation
• Performance results
• Conclusion and future work

24
Conclusion and Future Work

• Governor extensible framework for quantitative
  performance impact control
• Contains actual performance impact
• Proactively consume idle resources
• Self-adaptive
• OS-independent and low-overhead
• Future work
• Connect impact control with user interfaces
• Studying memory resource throttling
• Evaluating with more scavengers

25
Resource Utilization and ß for Composite