Building Communication Systems from Components Ensemble and Nuprl

1
Building Communication Systems from Components
(Ensemble and Nuprl)
2
Key observations

• VLSI uses component based approach to build
  larger systems
• Why cant we mimic that for software systems?
• Problems
• Layering inefficiencies
• Additional function calls
• Loss of locality
• Redundant code and unused header fields
• Separate compilation leading to poor optimization

3
Advantages of components

• Easy to design, test, optimize individual
  components
• More readily adapted and tuned for new
  environments than monolithic systems
• Facilitates evolution since easily extensible
  with additional components
• How can we get the advantages without
  sacrificing performance?
• Put theory and practice together

4
Cornell experiment

• Ensemble
• Protocol stack built using micro-protocols
• OCaml (see Caml origin) used as the programming
  language
• Formal semantics
• Efficient machine code
• Specification using IOA
• Syntax close to C
• Allows composition of subsystems specified using
  IOA
• Formal methods using Nuprl
• Logic system for program development,
  verification, and optimization
• Understands IOA spec and OCaml code
• Correctness can be verified
• Code can be optimized

5
Justification

• Ensemble has over 60 components
• Set of micro-protocols (sliding window,
  fragmentation/reassembly, flow control,
  encryption, )
• Can be composed to realize a variety of
  communication protocol stacks (eg. TCP/IP)
• Micro-protocol interface structure
• Top level to talk to layer above
• Bottom level to talk to layer below
• Event driven interface

6

• OCaml
• Leads to compact code
• High level ops on data structures
• Automatic memory allocation, GC, marshaling
• Programmability compared to C
• Precise definition leading verifiability of code
• IOA
• Gives high level logical spec of micro-protocol
  components
• Nuprl
• Tools convert OCaml code to Nuprl and vice versa
• Translated OCaml code can be reasoned and
  evaluated

7
Specification and Implementation

• Properties spec
• E.g. in order delivery
• Behavioral spec
• E.g. ack every packet
• Abstract behavior spec
• Simple, lends itself deriving properties (such as
  FIFOness)
• Not executable
• Concrete spec
• Close to implementation
• Implementation
• Small difference from concrete spec
• Order of execution of actions
• Condition checking for specific actions

8
Two abstract specs
FIFO network
Lossy network
9
Concrete Specification FIFO on lossy network
10
Putting them all together
11
Synthesizing stacks

• Given a behavior spec
• Can a stack (composed of Ensemble
  micro-protocols) be synthesized?
• Brute force will not work, why?
• Ensemble approach
• Algorithm that synthesizes a stack given set of
  properties (currently 24 props) and designers
  knowledge
• No guarantee that the stack is correct!
• while possible not yet automated
• Beware of bugs in the above code I have only
  proved it correct, not tried it!! -- Donald Knuth

12
Optimization opportunities in Ensemble

• Use internal explicit memory management than
  OCaml implicit GC
• Avoid marshaling and unmarshaling across layers
• Buffering in parallel with xmission
• Identifying and colocating common code path
  across layers for locality
• Header compression to eliminate redundant or
  unnecessary fields

13
Formal optimization using Nuprl

• Common Case Predicate (CCP)
• State of protocol and input event
• E.g. deliver event if sequence number of received
  packet equals that expected
• Synthesized from the conditional statements in a
  micro-protocol
• Generate bypass code if CCP true
• Can bypass several layers
• Can compress headers

14
Nuprl methodology

• Static level
• Depends on the micro-protocol code
• Semi-automatic with designer help
• Dynamic level
• Completely automated
• Composes layers
• Generates bypass code
• Optimization vs. verification?

15
Static optimization

• Layer by layer
• Code transformations
• Function inlining
• Directed equality substitutions
• Specific to functional programming leading to
  code simplicity
• Context dependent simplifications
• Helps in synthesizing code for the CCP bypass

16
Dynamic optimization

• Optimization across several layers
• Fully automated
• Optimization theorem
• Proves equivalence of CCP bypass code to the
  ensemble protocol stack
• Optimization theorem generated automatically
• Final step
• Results of all the optimization theorems
  converted to OCaml code
• Resultant program equivalent to original stack
  but more efficient

17
Big picture
18
Performance

• Four versions
• IMP imperative (normal) version
• FUNC functional composition version
• HAND hand optimized bypass code for special
  stacks
• MACH Nuprl methodology to generate this machine
  optimized stack

19
Results
300 MHZ UltraSparc 100 Mbps Ethernet
Sensitivity to message size
20
Connections to other work we have already seen

• Specializing is not new
• Done in SPIN, Exokernel, Thekkath and Levys work
• What are the similarities?
• What are the differences?