1. Potentially Interesting Startup and/or Commercialization Opportunities
Peter A. Dinda
Prescience Lab
Department of Electrical Engineering and Computer Science
Northwestern University

2. Outline Virtualization-based Computing
Virtuoso system and related software
Adaptive/autonomic computing
High performance computing
Desktop replacement computing
Performance analysis and prediction
RPS system
Advanced power management
User-driven and Process-driven DVFS
Hardware for low power sensor systems

3. Distributed Optical Testbed (DOT) Private Optical Network
IBM xSeries
virtual cluster
(64 CPUs),
1 TB RAID
Interactivity
Environment
Cluster, CAVE
(~90 CPUs),
8 TB RAID
Internet
2 Distributed
Optical Testbed
Clusters
IBM xSeries
(14-28 CPUs),
1 TB RAID
DOT clusters
with optical
connectivity
IBM xSeries
(14-28 CPUs),
1 TB RAID:
Argonne, U.Chicago,
IIT, NCSA, others
Nortel Optera
Metro Edge
Optical Router
Distributed Optical Testbed
(DOT) Private Optical Network
Northwestern

4. Users already know how to deal with this complexity at another level

5. Virtuoso: A Virtualized Computing Infrastructure
Providers sell computational and communication bandwidth
Users run collections of virtual machines (VMs) that are interconnected by overlay networks
Replacement for buying machines
That continuously adapts… to increase the performance of your existing, unmodified applications and operating systems
See web site for many papers, talks, and movies

6. The Illusion
User’s LAN VM
User
Your machines are sitting next to you.

7. Claim (initially published in 2003)
VMs interconnected with overlay networks enables the broad application of dream techniques…
Inference/measurement/prediction
Adaptation
Resource reservation
Marketplaces
… using existing, unmodified applications and operating systems
So actual people can use the techniques

8. Results Tools we have built In progress Virtuoso marketplace
VNET overlay network
VTTIF application topology inference
Wren interface for free network monitoring
VSCHED VM scheduling
VADAPT adaptation framework
VRESERVE optical network reservations
VTL transparent network services
In progress
VMM
VT-I/O transparent I/O services
Speculative remote display engine

9. Results Formalizations Algorithms / analysis Two Ph.D. theses so far
Adaptation problem with formal objective
With hardness and approximatability results
Adaptation problem with human-directed objective
Algorithms / analysis
Automatic inference of application topology and traffic matrix from existing traffic
Automatic inference of network bandwidth and latency from existing network traffic
Effective heuristic solutions for adaptation problems
Ad hoc, evolutionary, and human-directed
Significant speedup results for existing, unmodified HPC and web applications
Speculative remote display
Eliminating round-trip delays through prediction
Two Ph.D. theses so far
We have demonstrated that our claims are valid!

10. Results Numerous papers in high quality venues
Students / potential employees/partners
Ananth Sundararaj, Ph.D.
Adaptation formalization, heuristic algorithms, VNET, VADAPT, VRESERVE
Currently at Microsoft
Bin Lin, Ph.D.
Human-directed adaptation, VSched
Will soon join Intel
Dong Lu, Ph.D.
Scalable relational information services
Currently at Lehman Brothers
Jack Lange (Current Ph.D. student)
VNET, VRESERVE, VTL, Virtuoso marketplace
Ashish Gupta, (Current Ph.D. student)
VTTIF, VADAPT, Wren, etc. Currently exploring inference in greater depth for thesis
Joining Google in 2008
Alex Shoykhet, Blair Heuer, Sam Rossoff, Jay Bruins (former B.S. students)
Virtuoso Marketplace, Speculative display

11. Our Potential Niche… Adaptive/autonomic computing…
In a high performance computing context…
Each user owns multiple communicating VMs…
Which run parallel and high communication distributed applications
Which may be spread over the wide area, not just a single machine room
Unlike VMware, XenSource, etc.
and a desktop replacement context…
Where the desktop VM remains on the server side
Unlike Moka5, Intel ISR, etc.
Using existing, unmodified applications and OSes

12. Possible Business Models
License software to let others create / expand utility computing services
Amazon EC2, Sun Grid
Sell service on our own backend
“Your Desktop on the Web”
Pretend to be a hardware company, but one that charges per cycle instead of per box
Give software away, charge as a matchmaker between users and providers
Get bought…
VMWare, Moka5, IBM, Sun, …
License patents…

13. Our Prospects? Volume of HPC marketplace
Growing, esp. with availability of cheap multicore
But tends to be conservative
Information disclosure issue
For desktop replacement, it all hinges on where web applications (AJAX) are going and how much buy in there is
If AJAX gets better and people adopt AJAX apps, the need for a real desktop replacement declines. The browser becomes the thin client
Conceivably, we could compete in the virtual colocation space, but that seems crowded

14. Performance Prediction
Long chain of work, stretching back about 10 years
How can we cheaply measure and predict…
Supply side…
Host and net static characteristics
CPU load, availability
Network path and link bandwidth, latency
Serial and parallel TCP throughput
...
Demand side…
CPU, network, etc, demands
User interactions
User satisfaction …
To answer high-level questions…
Where should I send task X to get it done in 10 seconds?
If I send a 1 MB message to host H, how long will it take to arrive?
Find me a collection of 16 tighly connected machines with a total of 48 GB of RAM

15. The RPS Toolkit
Dinda’s RPS Toolkit facilitates the creation of online forecasting systems for resource supply and demand in distributed systems
Network bandwidth and latency, host load, disk bandwidths, etc…
Time series modeling
Better to write this in form the form of an API call

16. Niches / Models / Prospects
Unclear

17. Power Management
Modern processors can have their clock frequency and voltage set by software
Windows DVFS on laptops
Similar features on cell phones, PDAs, etc.
We have developed techniques that lead to dramatic power reduction (and thus longer battery lifetimes)
Combination of two ideas
User-driven Frequency Scaling
Get direct feedback about user satisfaction, set clock frequency just above the point of user annoyance
Process-driven Voltage Scaling
Find the minimum voltage for each frequency / temperature level
Customize to the individual user (lots of variation!) and customize to the individual machine (ditto)

18. User and Process Driven Dynamic Voltage and Frequency Scaling
Processors can dynamically change their clock frequency and voltage
Linear (frequency) and quadratic (voltage) effects on power consumption
Current algorithms do not customize to individual user or processor
We (Dinda, Memik, Dick) have developed the first algorithms to do so
User-driven Frequency Scaling (UDFS) customizes to user
Process-driven Voltage Scaling (PDVS) customizes to processor
User studies show dramatic reduction in power consumption
This is system power measurement on IBM T42 with screen turned off, replaying user traces from the FIFA game
% improvement over Windows DVFS
Users->

19. Niches / Models / Prospects
Niche: become the patent / software that everyone implements in their OS to get these dramatic power savings
Microsoft Windows, Apple MacOS, iPods, your cellphone, etc.
Models
Patent licensing
Patent + software licensing
Prospects
Seem very promising

20. Hardware for Low Power Sensor Systems
Long-lived sensor networks sorely needed
Low maintenance => year or more on a battery
Pull some of the sensor network processing out of software and implement as hardware

21. Archetype-Based Sensor Network Synthesis
Monitoring is a critical component of understanding, forecasting, and adapting to demands, including demand for power
Sensor networks promise to make monitoring ubiquitous…
…But they are incredibly difficult to develop, especially for real-world long-term, low-maintenance deployment
We have developed a new architecture and hardware for low-power event-driven applications, e.g., structural monitoring
250x power improvement
Custom hardware and software
Our long term goal is to make low-power sensor networks easy to design for application experts, facilitating ubiquity
Integrating language design, compiler, and hardware synthesis research with custom hardware design
Researchers: R. Dick, P. Dinda, C. Dowding (CEE), L. Henschen, S. Jevtic, M. Kotowsky (ITI)
Now NSF funded

22. Niches / Models / Prospects
Toolchains for non-experts
Hardware boards for low power “outboard” processing in existing sensor nodes
Models
License hardware designs or toolchain
Sell either/both directly
Consultingware
Prospects
Too early to tell about toolchain (research project)
Current hardware design is conceptually simple (devil is in the details) so how to protect it?
Too early to tell what to do with other hardware designs
Have “customer” (crack monitoring)
Have expertise to put together company, including graduates and other students

23. For More Information Peter Dinda Prescience Lab http://pdinda.org