Ioana Burcea Won-Ho Park

Electrical and Computer Engineering Department

University of Toronto

Algorithms for Implementation of Tuple Space

Expert TopicECE 1770 Spring 2003

Agenda

Ioana

- Matching Algorithms (in Linda)

Basic Terminology and Concepts

Multi-Key Search Algorithm

Won-Ho

- Linda: Implementation of Tuple Space

Basic Terminology

Tuple (out operation)- Ex:

(1, 4.0, “Middleware”) Template = antituple (in operation)

- Ex: (1, 4.0, ?x) x: string (1, 4.0, “Middleware”)

Tuple arity- Ex:

(1, ?x) arity 2 (1, 4.0, “Middleware”) arity 3

Type signature- Ex:

(1, 4.0, “Middleware”) (int, float, string)

Matching Algorithms

Exhaustive search – too time consuming Optimizations

- Segmenting Orthogonal subspaces

- Based on

» tuple arity & type signature» constants

- Treated separately

- Searching Zero variable case

- In(“mutex”) Out(“mutex”) One variable case

- Single key index Multiple variable

- Multiple key search

Patterns

Actual parameter- Ex:

(1, 4.0, “Middleware”) 1, 4.0, “Middleware” (1, 4.0, x) 1, 4.0, x

Formal parameter (placeholder / wildcard)- Ex:

?x (x: string) Tuple Pattern

- Bitvector its size is given by the arity of the tuple a bit is set if and only if it corresponds to a actual parameter

- Ex: (1, 4.0, “Middleware”) (111) (1, 4.0. ?x) (110)

Matching Semantics Using Patterns

Compatible patterns- Two tuples P and Q are compatible if and only if P OR Q

contains only 1s Two tuples may match if and only if their patterns are

compatible Must-match pattern

- Ex: Out(a, b, c) pattern (111) In (x, ?y, z) pattern (101)

Matching definition- Two tuples match if and only if

They have the same type signature Exactly one is an antituple Their patterns are compatible Their must-match values are equal

OR = (111) compatible AND = (101) must-match pattern

Multiple key search Applied inside each orthogonal subspace Basic idea:

- Build the key search based on the pairs of compatible patterns and must-match values

Ex:- Tuple pattern: (111)

- Antituple patterns: (010), (100), (110) an entry in the dictionary for each pair (tupple pattern, antituple

pattern) => each tuple will be referenced by multiple keys the key of the entry also contains the must-match values the entry points to the tuple Ex: out(3, 5, x) => Key: (010:111:5)

- ! One record per must-match pattern is not sufficient ! Property:

- If two tuples have the same key => they match

Perfect hash function – one to one correspondence between search keys and the hash keys

Offline Algorithm – in operation1) Lock subspace2) Identify antituple’s pattern, p, and store antituple

where it can be found via a pointer or other reference.3) For each compatible tuple pattern, c, in use,4) Combine p, c, and antituple’s must-match values

to form a search key, k.5) Search dictionary for a tuple reference indexed by k.6) If one is found, then7) Delete tuple, antituple and all their references.8) Goto 10. else9) Insert a reference to antituple, indexed by k, into dictionary.10) Unlock subspace11) If no match was found, then 12) Suspend execution until signalled else13) Return match result

Online Algorithm

Adapt the tuple space whenever a new pattern is encountered

2.1) If p is a new antituple pattern, then2.2) Create a new list for p2.3) For each old, compatible tuple pattern q2.4) For each tuple t on q’s list2.5) Insert into dictionary a new reference to t, based on p and q2.6) Insert antituple in p’s list

Discussion- Usually, no perfect hash

Each hash entry has a hash chain The hash chain has to be checked until a match is found

Virtual Linda Machine (VLM):Implementation of Linda on Networks

TupleSpaces is a virtually shared memory: - How to implement TupleSpaces in the absence of shared

memory

- How to find Tuples and where keep them Components of VLM

- Hardware

- Network Communication Kernel

- Compiler Hardware is networks, or network computers The kernel is the program resident on each network node

that implements inter-node communication- Performing In(), out(), read() primitive instructions on

networks by sending or broadcasting messages The compiler adds code for maintaining buffer (shared

memory)

Run-Time Name Resolution in VLM:How to find Tuples

VLM should provide Name(Tuple)-to-Node(address) mapping information

- Centralized in directory nodes

- Distributed network-wide (e.g., by means of a network-wide broadcast)

The communication kernel translates Tuple-names into network-addresses on a per-reference basis

Network state and its storage

- Only one node generating state information maintains the state

- All nodes maintain all state by broadcasting

- VLM uses N1/2-Node Broadcast Technique.

N1/2-Node Broadcast Techniqueand Name Resolution

Write set : row nodes Read set : column nodes P-in-thread process P-out-thread process

P-in-thread is intercepted by a P-out-thread on node k.

A notification message is generated on k-node, and sent to m-node

The message instructs m-node to send tuple(P…) to n-node.

mm

P-out-P-out-threadthread

Out(P…)Out(P…)

In(P…In(P…))

kk nn

P-in-threadP-in-thread

A Network Topology

Buffering:Where to keep Tuples

VLM is required to simulate Linda’s infinite global buffer Buffer is allocated by the compiler-generated code, not by

the kernel

- in(P….) : allocating a buffer on the execution stack of the process, and referred to buffer P.

- out(P…) : Tuples are stored in per-process heaps, and destroyed when it is delivered

- Header information required by the communication system If a process terminates with undelivered Tuples

- The buffers are maintained by the communication kernel until delivery occurs