Transcript Examples of OS Research that are Driven by

Your Speaker in a Nutshell
BA in Economics 1965
PhD University of CA, Berkeley 1989
Technical skills and innovations that married architecture & compilers.
Mostly collaborations.
Pushed for quality over quantity, technology transfer & impact
I have always loved my job.
Thesis: better understanding of sharing behavior
MVP: Simultaneous Multithreading
Current work: a new life for dataflow machines
Transitioning to retirement (golf, gardening & landscaping, food snob).
October 2007
Susan Eggers: SOSP Women's
Workshop
1
Challenge in Computer Architecture
There’s one biggie: programs run faster
October 2007
Susan Eggers: SOSP Women's
Workshop
2
The Underlying Cause
October 2007
Susan Eggers: SOSP Women's
Workshop
3
How Architects are Addressing the Challenge
Some things we have some idea how to do:
• More aggressive speculation (threads, memory, value)
• More aggressive recovery mechanisms (transactional memory)
• Lighter-weight locking mechanisms
• Exploiting super-fine-grain parallelism with dataflow execution
• Adding a few more processors
• The engineering: building hardware for sharing and concurrency
October 2007
Susan Eggers: SOSP Women's
Workshop
4
How Architects are Addressing the Challenge
Some things we have not a clue about:
• Writing parallel codes for 100s or 1000s of processors
• Discovering other than embarrassingly parallel, coarse-grain
parallelism
• Determining the best programming model for writing parallel
codes
• Should the programmer detect parallelism or should the
compiler do it automatically?
• Devising language constructs to express parallelism
• Exploiting the concurrency once we’ve discovered it
• Avoiding race conditions, deadlock, livelock
• Doing all this in a way that is simple and intuitive for the
programmer
October 2007
Susan Eggers: SOSP Women's
Workshop
5
How Architects are Addressing the Challenge
Some things we have not a clue about:
• Executing with 100s or 1000s of processors
• Making good trade-offs between load balancing & locality
• Recovering from failing processors (RAID issues)
October 2007
Susan Eggers: SOSP Women's
Workshop
6
Rules of Thumb for Picking New Research Projects
How to develop a nose for good research ideas.
What in the underlying technology or application space has changed?
• How has it changed?
How does that affect my area of interest?
• In what areas can I no longer do business as usual?
• Where are the new performance bottlenecks?
• What opportunities does the change bring?
• What technology or techniques that went out of style can be
resurrected?
Caveat: technology trends don’t always turn out the way you expect
• CCDs failed in computers; Prologue failed
• Are supercomputers failing?
• Will quantum and DNA
computing succeed?
October 2007
Susan Eggers: SOSP Women's
Workshop
7
Examples of OS Research that are Driven by
Computer Architecture
The operating system:
• How do you design an OS for 1000s of processors?
• How do current OSs hinder scalability? Where are the
bottlenecks? Where are the centralized software structures?
• Do you start from scratch?
• Great opportunity to design free of tradition & an industry
standard
• What should be the key design criteria?
• Do you use virtual machines to execute programs on small clusters
of processors and then build a programming environment that hides
that?
October 2007
Susan Eggers: SOSP Women's
Workshop
8
Examples of OS Research that are Driven by
Computer Architecture
Transactional memory:
• How should programs written with transactions interface with the
OS?
• What does rollback mean over a network?
Security:
• What if everyone is a parallel programmer?
• latent concurrency bugs, deadlock, livelock
• What are the implications for security?
October 2007
Susan Eggers: SOSP Women's
Workshop
9
Examples of OS Research that are Driven by
Computer Architecture
Fine-grain threads:
• The OS has to support more than pthreads.
• How fine-grain can/should they be?
• And stepping back, what is a thread? Can it be data centric?
Virtual machine support in hardware:
• What if virtual machines were very light-weight?
• What might you use them for and how?
• Can a virtual machine replace a process? a protection domain?
October 2007
Susan Eggers: SOSP Women's
Workshop
10
Examples of OS Research that are Driven by
Something Else
New applications:
• How will Google’s needs change the OS?
• How should the OS manage fast I/O between CPUs (10Gb)?
• How can the OS make up for its absence?
Phase-change memory (fast, non-volatile, but high power)
How would you build a file system if you didn’t need disks?
Third-world computing:
• Low-cost, low-power, multi-lingual, multiple education levels,
different apps
• What does the OS look like?
More powerful processor, same number of pins?
• How do you feed it?
October 2007
Susan Eggers: SOSP Women's
Workshop
11