1

Multithreaded Architectures

Lecture 3 of 4

Supercomputing ’93 Tutorial

Friday, November 19, 1993

Portland, Oregon

Rishiyur S. Nikhil

Digital Equipment

Corporation

Cambridge Research Laboratory

Copyright (c) 1992, 1993 Digital Equipment Corporation

nikhil−su93−3

Supercomputing '93 Multithreaded Architectures Tutorial, Lecture 1 2

Overview of Lecture 3

Multithreading: a Dataflow Story

Dataflow graphs asa machine language

MIT Tagged Token ManchesterDataflow Architecture Dataflow

ETL Sigma−1

Explicit TokenStore Machines

MIT/Motorola ETL EM−4Monsoon

P−RISC: "RISC−ified" dataflow

MIT/Motorola *T

nikhil−94

Supercomputing '93 Multithreaded Architectures Tutorial, Lecture 1 3

Dataflow Graphs as a Machine Language

Code:Dataflow Graph(interconnectionof instructions)

Conceptually, "tokens" carryingvalues flow along the edges ofthe graph.

Memories:Two major components

Heap, for data structures(e.g., I−structure memory)

"Stack"(contexts, frames, ...)

Values on tokens may be memoryaddresses

Each instruction:

Waits for tokens on all inputs(and possibly a memorysynchronization condition)

Consumes input tokens

Computes output values basedon input values

Possibly reads/writes memory

Produces tokens on outputs

No further restriction on instructionordering

A common misconception:

"Dataflow graphs and machinescannot express side effects,

and they only implementfunctional languages"

nikhil−95

Supercomputing '93 Multithreaded Architectures Tutorial, Lecture 1 4

Encoding Dataflow Graphs and Tokens

Conceptual Encoding of graphProgram memory:

Op−

op1 op2

+

op3 op4

Encoding of token:A "packet" containing:

code

109 op1

113 op2

120 +

141 op3

159 op4

Destination(s)

120L

120R

141, 159L

... ,

120R Destination instruction address, Left/Right port6.847 Value

Re−entrancy ("dynamic" dataflow):

Each invocation of a function or loop iteration getsits own, unique, "Context"

Tokens destined for same instruction in differentinvocations are distinguished by a context identifier

120R Destination instruction address, Left/Right portCtxt Context Identifier6.847 Value

nikhil−96

Supercomputing '93 Multithreaded Architectures Tutorial, Lecture 1 5

MIT Tagged Token Dataflow Architecture

Designed by Arvind et. al., MIT, early 1980’s

Simulated; never built

Global view:

Processor Nodes

(including local program and "stack" memory)

ResourceInterconnection Network ManagerNodes

"I−Structure" Memory Nodes (global heap data memory)

Resource Manager Nodes responsible for

Function allocation (allocation of context

identifiers)

Heap allocation etc.

Stack memory and heap memory: globally addressed

nikhil−97

Supercomputing '93 Multithreaded Architectures Tutorial, Lecture 1 6

MIT Tagged Token Dataflow ArchitectureProcessor

Wait−Match Waiting tokenUnit memory (associative)

Instruction ProgramFetch memory

TokenQueue

ExecuteOp

FormTokens

Output

To network From network

Wait−Match Unit:

Tokens for unary ops go straight through

Tokens for binary ops: try to match incoming tokenand a waiting token with same instruction addressand context idSuccess:Both tokens forwarded

Fail: Incoming token −−> Waiting Token Mem,Bubble (no−op) forwarded

nikhil−98

Supercomputing '93 Multithreaded Architectures Tutorial, Lecture 1 7

MIT Tagged Token Dataflow ArchitectureProcessor Operation

120R,c, 6.847Waiting token

Wait−MatchUnit

120,c, 6.001,6.847

InstructionFetch

Token 141,159L,c, +,6.001,6.847Queue

ExecuteOp

141,159L,c, 12.848

FormTokens

141,c, 12.848

159L,c, 12.848

Output

memory (associative)120L,c, 6.001

Programmemory

120 + 141, 159L

To network From network

Output Unit routes tokens:

Back to local Token

Queue

To another Processor

To heap memorybased on the addresses on the token

Tokens from network are placed in Token Queue

nikhil−99