Hints for Computer System Design

1
Hints for Computer System Design

• Butler W. Lampson

2
Butler Lampson - Background

• Founding member of Xerox PARC (1970), DEC
  (1980s), MSR (current)
• ACM Turing Award (1992)
• Laser printer design
• PC Alto
• Two-phase commit protocols
• Bravo, the first WYSIWYG text formatting program
• Ethernet, the first high-speed LAN

3
System Design Introduction

• complex act
• less precisely defined, changing requirements
• Choices
• Affects future choices
• Affects system-wide performance
• But how?

4
System Design Introduction

• Other reasons its so hard
• Moores law
• OSs dont improve by a factor of 100 every decade
• Backward compatibility
• OSs are fundamentally different than standard
  apps.
• Extremely large programs
• UNIX 30 Million lines of code
• Windows Vista 50 million lines of code
• OSs have to deal with potentially hostile users
  at same time allow users to share information

5
System Design Introduction

• Even more reasons its so hard
• OSs live for a very long time
• UNIX is over a quarter of a century old
• Must prepare for hardware/software changes
• OS designers do not have a good idea how their
  systems will ultimately be used
• Email web browsers came after UNIX
• OSs have to run on multiple hardware platforms
• OSs frequently have to be backward compatible

6
System Design The Slogans
7
Functionality The Interface

• Interface separates implementation of some
  abstraction from clients using the abstraction
• Eg File (open, read, write, close)
• Desirable properties
• Simple
• Complete
• Admit small and fast implementation

8
Keep It Simple

• Do one thing at a time, and do it well
• Example PL/1compiler
• Early implementation is slow and inefficient
• For compiler optimization, it takes 15 years

9
Keep It Simple

• too much large, slow and complicated
  implementation
• Should have predictable (reasonable) cost.
• Avoid features needed by only a few clients

10
Example

• Tenex System
• reference to an unassigned page -gt trap to user
  program
• arguments to sys calls passed by reference
• CONNECT(string passwd) -gt if passwd wrong, fails
  after a 3 second delay
• CONNECT
• for i 0 to Length(directoryPassword) do
• if directoryPasswordi ! passwordArgumenti
  then
• Wait three seconds return BadPassword
• end if
• end loop
• connect to directory return Success

11
Breaking CONNECT(string passwd)
Unassigned Page
Assigned Page
A
B
Bad Passwd
Invalid page
12
Breaking CONNECT(string passwd)
Worst case 64n tries as opposed to 128n/2
tries n passwd length (bytes)
Unassigned Page
Assigned Page
B
A
Z
Bad Passwd
Invalid page
13
Functionality The Interface

• Get it right
• eg FindIthField O(n), FindNamedfield O(n2)
• Make it fast rather than general or powerful
• RISC machine
• Dont hide power
• Virtualization virtual hardware exposed is
  functionally identical to the underlying hardware
• Exokernel the high-performance of the hardware
  is just exported to the user

14
Functionality The Interface

• Use procedure arguments to provide flexibility in
  an interface
• Leave it to the client
• Monitors locking signaling mechanisms do very
  little, leaving all the real work to the client
• Keep basic interfaces stable
• Keep a place to stand if you must change
  interfaces (backward compatibility)
• Emulation library microkernels emulate
  interfaces by trap redirection
• Virtualization VMM multiplexing at the
  granularity of an operating system

15
Functionality Implementation

• Plan to throw one away
• Keep secrets of the implementation
• Virtualization
• Hide the messy stuff like NUMA
• Monitors
• Allow access only at certain entry points
• Use a good idea again instead of generalizing it
• Divide and Conquer
• Dover laser printer
• Deal with 33004200 bits (14 M bits)

16
Functionality Completeness

• Handle normal and worst case separately
• The normal case must be fast
• The worst case must make progress
• Caches and hints
• Sometimes radically different strategies are
  appropriate in the normal and worst cases.
• One crash a week is usually a cheap price to pay
  for 20 better performance.

17
Speed Interface

• Split resources in a fixed way if in doubt rather
  than sharing them
• I/O channel. Floating-point coprocessors
• Use static analysis if you can
• Discover properties of the program that can
  usually be used to improve its performance
• Dynamic Translation
• Smalltalk uses the bytecodes and a cache to store
  translated code.

18
Speed Implementation

• Cache answers to expensive computations
• Use hints
• The Ethernet exponential backoff
• Links in web pages
• When in doubt use brute force
• Compute in background
• Use batch processing
• Cleanup in log structured file systems segment
  cleaning could be scheduled at nighttime.

19
Speed Completeness

• Shed load to control demand
• SEDA (Staged Event Driven Architecture)
• Adaptive load shedding
• Apache
• Bounded thread pool
• Safety First
• When allocating resources remember safety first
• Avoid temptation to try to obtain optimum
  performance (instead try to avoid disaster)
• Progression of kernel structuring approaches for
  extensibility
• Microkernels, sandboxing, SPIN vs. Mach (some
  versions) Chorus

20
Fault-tolerance

• End-to-end
• Error recovery at the application level is
  absolutely necessary
• Log updates
• Logs can be reliably written/read
• Logs can be cheaply forced out to disk, which can
  survive a crash
• Log structured file systems
• Make actions atomic or restartable
• Transaction processing system
• Database systems have provided atomicity

21
System Design Conclusion

• Just remember
• Designing an OS starts with determining what it
  should do
• The interface should be simple, complete,
  efficient
• There probably isnt a best solution for even
  one part of any system
• Just dont choose a terrible solution
• Have a clear division of responsibility among the
  parts

22
System Design The Slogans
23
System Design Quotes

• Perfection is reached not when there is no
  longer anything to add, but when there is no
  longer anything to take away
• --A. Saint-Exupery
• The unavoidable price of reliability is
  simplicity.
• --C. Hoare

24
System Design References

• Jon Walpole
• Hints for Computer System Design Lampson.
• Modern Operating Systems, Second Edition.
  Tannenbaum, pp. 855-894.
• http//c2.com/cgi/wiki?SpiralModel
• http//www.dpw.wau.nl/pv/temp/clipart/
• http//firstmonday.org/issues/issue7_11/tuomi/