The MTEvaluator Virtual Machine

1
The MT-Evaluator Virtual Machine

• Kristine Joy Apon, Kai H. Fan, Joanie Feeney,
  Peter Laurina, Steve OBrien, Gregory Van De
  Moere
• Advisor Dr. Marco Morazán

2
Functional Programming

• Describing processes such as mathematical
  notations and procedures
• (define (fact n)
• (if ( n 0)
• 1
• ( n (fact (- n 1))) ) )

3
Functional Programming

• Benefits
• More intuitive than imperative languages such as
  C and Java
• Easy to prove correctness of functions
• Truly machine independent
• Fast prototyping

4
Disadvantages

• Poor virtual memory interaction can make them
  slow.
• Excessive paging can profoundly affect
  performance.

5
Distributed Virtual Memory System

• Tailored for functional languages implemented
  completely in software.
• Parallelize the engine that evaluates programs
  and not individual user code.
• Parallelize memory management.

6
MT Architecture
7
Virtual Machines

• Software that emulates a real machine by keeping
  track of state.
• Can port software from one machine to another.
• Can execute software designed for a machine never
  physically implemented.
• The MT evaluator is a switch-based virtual
  machine.

8
MT Evaluator Virtual Machine
9
Implementation

• Heap, stack, symbol table, and code are objects
  implemented in C
• Instance variables
• For data
• For virtual memory implementation
• Public and private methods
• Public methods constructor, data access methods.
• Private Methods memory management routines.
• Paging occurs within these discretized memory
  spaces.
• No shared memory (either pages or frames) between
  spaces.

10
Paging Policy Considerations

• Paging behavior of distributed spaces
• Heap
• FIFO LRU (FIFO has less overhead).
• Stack
• LRU superior to FIFO.
• Uses a special version of LRU without timestamps.
• Code
• DLL-LRU LRU employing a circular double-linked
  list.

11
Test 1 Creating a list of 10000 random numbers
using FIFO

• Empirical Measurements for performance
• Heap
• 40000 accesses
• 20000 allocations
• 60 faults (.0015 fault rate)
• Stack
• 270013 accesses
• 120006 allocations
• 99 faults (.0004 fault rate)

12
Future Work

• Comparing FIFO LRU, other paging algorithms for
  code space and garbage collection.
• Implementation of 1st-class functions.
• Development of an MT Compiler for different
  functional languages.
• Design and implementation of a GC algorithm that
  exploits the MT architecture.