additional slides - MIT

nikhil−96
Encoding Dataflow Graphs and Tokens
Conceptual Encoding of graph
op1 op2
+
op3 op4
Encoding of token:
A "packet" containing:
109
113
120
141
159
Program memory:
op1
op2
+
op3
op4
Destination(s)
120R Destination instruction address, Left/Right port
6.847 Value
Re−entrancy ("dynamic" dataflow):
Op−
code
120L
120R
141, 159L
... , ...
Each invocation of a function or loop iteration gets
its own, unique, "Context"
Tokens destined for same instruction in different
invocations are distinguished by a context identifier
120R Destination instruction address, Left/Right port
Ctxt Context Identifier
6.847 Value
...

MIT Tagged Token Dataflow Architecture
nikhil−97
Designed by Arvind et. al., MIT, early 1980’s
Simulated; never built
Global view:
Processor Nodes
(including local program and "stack" memory)
Interconnection Network
"I−Structure" Memory Nodes (global heap data memory)
Resource Manager Nodes responsible for
Resource
Manager
Nodes
Function allocation (allocation of context identifiers)
Heap allocation
etc.
Stack memory and heap memory: globally addressed

MIT Tagged Token Dataflow Architecture
Processor
nikhil−98
Token
Queue
To network
Wait−Match Unit:
Wait−Match
Unit
Instruction
Fetch
Execute
Op
Form
Tokens
From network
Waiting token
memory (associative)
Program
memory
Tokens for unary ops go straight through
Tokens for binary ops: try to match incoming token
and a waiting token with same instruction address
and context id
Success:
Fail:
Output
Both tokens forwarded
Incoming token −−> Waiting Token Mem,
Bubble (no−op) forwarded

MIT Tagged Token Dataflow Architecture
Processor Operation
nikhil−99
Token
Queue
120R,c, 6.847
Wait−Match
Unit
Instruction
Fetch
141,159L,c, +,6.001,6.847
To network
120,c, 6.001,6.847
Execute
Op
141,159L,c, 12.848
Form
Tokens
141,c, 12.848
159L,c, 12.848
Output
Output Unit routes tokens:
From network
Back to local Token Queue
To another Processor
To heap memory
based on the addresses on the token
Waiting token
memory (associative)
120L,c, 6.001
Program
memory
120 + 141, 159L
Tokens from network are placed in Token Queue

MIT Tagged Token Dataflow Architecture
Support for "Remote Loads"
nikhil−100
Conceptual:
A
Fetch
v
op1
Encoding of graph:
Program memory:
200
Opcode Destination(s)
Fetch
207 op1
Heap Memory
207
...
A v

MIT Tagged Token Dataflow Architecture
Support for "Remote Loads"
nikhil−101
Token
Queue
207,c, v
200,c, A
Fetch, A, 207,c
Fetch, A, 207,c
Wait−Match
Unit
200,c, A
Instruction
Fetch
207,c, Fetch,A
Execute
Op
Fetch,A, 207,c
Form
Tokens
Fetch, A, 207,c
Output
Network
Heap Memory Module
Containing Address A
207,c, v
207,c, v
A v
Multiple remote loads are no problem:
Waiting token
memory (associative)
Program
memory
200 Fetch 207
Can be issued in parallel
"Join" of responses is implicit in Wait−Match

MIT Tagged Token Dataflow Architecture
Support for "Synchronizing Loads"
nikhil−102
Heap memory locations have FULL/EMPTY bits
When "I−Fetch" request arrives (instead of "Fetch"),
if the heap location is EMPTY, heap memory module
queues request at that location
Later, when "I−Store" arrives, pending requests are
discharged
"I−structure semantics"
Note: no busy waiting, no extra messages
I−Store,A,v, 191,c2
I−Fetch,A, 207,c
I−Fetch,A, 900,c1
Heap Memory Module
Containing Address A
207,c, v
A E Nil A E
A F v
207,c
900,c1
191,c2, ack
900,c1, v

nikhil−180
Unification of Dataflow & von Neumann
Designs
pure Dataflow
Static
(Dennis &
Misunas ’74)
Argument
Fetching
(Gao,
et. al. 88)
"Static"
(Dennis 91)
Manchester
Dataflow
(Gurd &
Watson 82)
Sigma−1
(Shimada &
Hiraki 87)
Epsilon
(Grafe
et.al. 88)
Epsilon 2
(Grafe &
Hoch 89)
EM 5
(Kodama
& Sakai 91)
TTDA
(Arvind
et.al. 80−83)
Monsoon
(Papadopoulos
& Culler 88)
EM 4
(Sakai 89)
Monsoon
Threads
(Traub &
Papadopoulos
91)
Unified
Hybrid
(Buehrer &
Ekanadham 87)
VNDF
Hybrid
(Iannucci 88)
P−RISC
(Nikhil &
Arvind 88)
Hybrid
*T
multithreaded
von Neumann
TAM
CDC 6600
MASA
(Halstead &
Fujita 88)
EMPIRE
(Iannucci
et. al. 89−91)
(Culler,
et. al. 90)
(Nikhil,
Papadopoulos,
Arvind &
Greiner 91)
1964
HEP
(B.Smith 78)
iPSCs,
CM−5
Horizon
Tera
(B.Smith,
et.al. 88−93)
Alewife
(Agarwal 89)
Message
Passing
Cosmic
Cube
(Seitz 85)
J−Machine
(Dally 87)
(adapted
from a slide
by Greg
Papadopoulos)