Integrated Scientific Workflow
Management for the Emulab
Network Testbed
Eric Eide, Leigh Stoller,
Tim Stack, Juliana Freire,
and Jay Lepreau
University of Utah,
School of Computing
USENIX 2006 / June 3, 2006
This Talk in One Slide
z
Current network testbeds
z
z
…manage the “laboratory”
…not the experimentation process.
z
→ A big problem for large-scale activities!
z
Evolve Emulab for experiments based on
scientific workflows
z
z
Big mutual benefits: testbed ↔ workflow
Work in progress
2
Example: UAV Simulation
z
A distributed, real-time
application
images →
z
Evaluate improvements
to real-time middleware
z
z
z
z
z
vs. CPU load
vs. network load
4 research groups
x 19 experiments
x 56 metrics
alerts →
UAV
Receiver
ATR
← images
3
Use Emulab
write “ns” file
Concept
Experiment
“swap in”
Emulate
4
Problems Solved
z
I get machines!
328 PCs, and more
z Time- & space-shared
z Loads OS and software
z
z
I get network!
z
z
Config. topology & quality
I get to collaborate!
Available to researchers
and educators worldwide
z File storage, email, …
z
5
Problems Not Solved
z
“Now what?”
z
Getting off the ground
Run all my software
z Add instrumentation
z Collect all my data
z Analyze it
z
z
Scaling up
19 configurations
z Automation
z
6
More Problems Not Solved
z
“How did I get here?”
z
Over the short term…
“Where are the results
I got last week?”
z “How did I get those
results anyway?”
z “What if…?”
z
z
…and the long term
Reproducing results
z Reusing artifacts
z
7
Idea: Scientific Workflow
z
Managing activities, inputs, and outputs is the
job of a scientific workflow system
z
Our approach: evolve Emulab with
integrated support for scientific workflows
z
z
z
Build on existing abstractions & mechanisms
Resource focus → user & task focus
Users work “within” and “across” experiments
8
Contributions
z
Address demand + opportunity
z
z
z
Advance the applicability of testbeds
z
z
Users need to manage large-scale complexity
A symbiotic combination: leverage and impact
Not just Emulab — e.g., PlanetLab and DETER
Advance scientific workflow systems
z
z
Exploit testbed capabilities — e.g., “total control”
Address testbed requirements — e.g., flexible use
9
Issue: Encapsulation
z
Current “experiment” model
is not fully encapsulating
z
z
z
Topology + static events
Need everything else!
ns file
packages
inputs
outputs
Challenge: specification
z
z
Complete and precise…
…w/o huge user burden
my software
z
OSes
Approach: be automatic
z
z
z
E.g., track files used
Snapshot, archive, restore
User can refine “extent”
NFS monitors
Subversion repo.
packet monitors
datapository (DB)
AJAX GUI
research filesystems
10
Issue: Definition vs. Execution
z
Current “experiment” has
multiple roles
z
z
z
Challenge: representing
relationships
z
z
z
Definition
The thing that you run
Multiple runs of one setup
Similar configurations
Approach: a new model of
experimentation
z
z
Separate the roles
Evolve the new abstractions
11
New Model
z
Template
z
Swapin
z
Experiment
z
Activity
z
Record
n=2
n=4
12
Issue: History
rev 1.1
z
Research and educational
plans are dynamic
z By design & by discovery
z
Challenge: safe exploration
z Fork
z Back up
z
Approach: keep history &
support temporal navigation
z Keep template revisions
z Track provenance
z Locate, repeat, and reuse
oops: need new
measurements
bigger nets
what about
loss?
add params
13
Implementation in Progress
Definition
Data Analysis
Execution
& History
14
Conclusion
z
Large and powerful testbeds
z
z
z
Integrated workflow management can
leverage the strengths of testbeds
z
z
…enable complex and large-scale activities
…lead to complex and large-scale workflow
management problems
Systems approach — and systems challenges
→ Better testbeds and workflow systems
15
http://www.emulab.net/
Thanks!
Extra Slides After This Point
17