1. 1 Chapter 2 Dataflow Processors

2. 2 Dataflow processors Recall from basic processor pipelining: Hazards limit performance. – Structural hazards – Data hazards due to true dependences or name (false) dependences: anti and output dependences – Control hazards Name dependences can be removed by: – compiler (register) renaming – renaming hardware  advanced superscalars – single-assignment rule  dataflow computers Data hazards due to true dependences and control hazards can be avoided if succeeding instructions in the pipeline stem from different contexts  dataflow computers, multithreaded processors

3. 3 Dataflow vs. control-flow von Neumann or control flow computing model: – a program is a series of addressable instructions, each of which either specifies an operation along with memory locations of the operands or it specifies (un)conditional transfer of control to some other instruction. – Essentially: the next instruction to be executed depends on what happened during the execution of the current instruction. – The next instruction to be executed is pointed to and triggered by the PC. – The instruction is executed even if some of its operands are not available yet (e.g. uninitialized). Dataflow model: the execution is driven only by the availability of operand! – no PC and global updateable store – the two features of von Neumann model that become bottlenecks in exploiting parallelism are missing

4. 4 Dataflow model of computation Enabling rule: An instruction is enabled (i.e. executable) if all operands are available. – Von Neumann model: an instruction is enabled if it is pointed to by PC. The computational rule or firing rule, specifies when an enabled instruction is actually executed. Basic instruction firing rule: An instruction is fired (i.e. executed) when it becomes enabled. – The effect of firing an instruction is the consumption of its input data (operands) and generation of output data (results). – Where are the structural hazards? Answer: ignored in traditional dataflow literature!!

5. 5 Dataflow languages Main characteristic: The single-assignment rule: A variable may appear on the left side of an assignment only once within the area of the program in which it is active. Examples: VAL, Id, LUCID A dataflow program is compiled into a dataflow graph which is a directed graph consisting of named nodes, which represent instructions, and arcs, which represent data dependences among instructions. – The dataflow graph is similar to a dependence graph used in intermediate representations of compilers. During the execution of the program, data propagate along the arcs in data packets, called tokens. This flow of tokens enables some of the nodes (instructions) and fires them.

6. 6 Dataflow architectures - Overview Pure dataflow computers: – static, – dynamic, – and the explicit token store architecture. Hybrid dataflow computers: – Augmenting the dataflow computation model with control-flow mechanisms, such as RISC approach, complex machine operations, multi-threading, large-grain computation, etc.

7. 7 Pure dataflow A dataflow computer executes a program by receiving, processing and sending out tokens, each containing some data and a tag. Dependences between instructions are translated into tag matching and tag transformation. Processing starts when a set of matched tokens arrives at the execution unit. The instruction which has to be fetched from the instruction store (according to the tag information) contains information about – what to do with data – and how to transform the tags. The matching unit and the execution unit are connected by an asynchronous pipeline, with queues added between the stages. Some form of associative memory is required to support token matching. – a real memory with associative access, – a simulated memory based on hashing, – or a direct matched memory.

8. 8 Static dataflow A dataflow graph is represented as a collection of activity templates, each containing: – the opcode of the represented instruction, – operand slots for holding operand values, – and destination address fields, referring to the operand slots in sub- sequent activity templates that need to receive the result value. Each token consists only of a value and a destination address.

9. 9 Dataflow graph and activity template

10. 10 Acknowledgement signals Notice, that different tokens destined for the same destination cannot be distinguished. Static dataflow approach allows at most one token on any one arc. Extending the basic firing rule as follows: An enabled node is fired if there is no token on any of its output arcs. Implementation of the restriction by acknowledge signals (additional tokens ), traveling along additional arcs from consuming to producing nodes. Using acknowledgement signals, the firing rule can be changed to its original form: A node is fired at the moment when it becomes enabled. Again: structural hazards are ignored assuming unlimited resources!

11. 11 MIT Static Dataflow Machine

12. 12 Deficiencies of static dataflow Consecutive iterations of a loop can only be pipelined. Due to acknowledgment tokens, the token traffic is doubled. Lack of support for programming constructs that are essential to modern programming language – no procedure calls, – no recursion. Advantage: simple model

13. 13 Dynamic dataflow Each loop iteration or subprogram invocation should be able to execute in parallel as a separate instance of a reentrant subgraph. The replication is only conceptual. Each token has a tag: – address of the instruction for which the particular data value is destined – and context information Each arc can be viewed as a bag that may contain an arbitrary number of tokens with different tags. The enabling and firing rule is now: A node is enabled and fired as soon as tokens with identical tags are present on all input arcs. Structural hazards ignored!

14. 14 The U-interpreter (U = unraveling) Each token consists of an activity name and data – the activity name comprises the tag. the tag has – an instruction address n, – the context field c that uniquely identifies the context in which the instruction is to be invoked, – and the initiation number i that identifies the loop iteration in which this activity occurs. Note, that c is itself an activity name. Since the destination instruction may require more than one input, each token also carries the number of its destination port p. We represent a token by

15. 15 The U-interpreter If the node n i performs a dyadic function f, and if the port p of n j is the destination of n i, then we have:

16. 16 MERGE and SWITCH nodes

17. 17 Branch Implementations Branch Speculative branch evaluation

18. 18 Basic loop implementation X f TF L D D L new x n i n j n l n m n k SWITCH P L: initiation, new loop context D: increments loop iteration number D -1 : reset loop iteration number to 1 L -1 : restore original context

19. 19 Function application A: create new context BEGIN: replicate tokens for each fork END: return results, unstack return address A -1 : replicate output for successors

20. 20 I-structures (I = incremental) Problem: Single-assignment rule and complex data structures – each update of a data structure consumes the structure and the value producing a new data structure. – awkward or even impossible to implement. Solution: concept of I-structure: – a data repository obeying the single-assignment rule – each element of the I-structure may be written only once but it may be read any number of times The basic idea is to associate with each element status bits and a queue of deferred reads.

21. 21 I-structures The status of each element of the I-structure can be: – present: the element can be read but not written, – absent: a read request has to be deferred but a write operation into this element is allowed, – waiting: at least one read request of the element has been deferred.

22. 22 MIT Tagged-Token Dataflow Architecture

23. 23 Manchester Dataflow Machine

24. 24 Advantages and deficiencies of dynamic dataflow Major advantage: better performance (compared with static) because it allows multiple tokens on each arc thereby unfolding more parallelism. Problems: – efficient implementation of the matching unit that collects tokens with matching tags. Associative memory would be ideal. Unfortunately, it is not cost-effective since the amount of memory needed to store tokens waiting for a match tends to be very large. All existing machines use some form of hashing techniques. – bad single thread performance (when not enough workload is present) – dyadic instructions lead to pipeline bubbles when first operand tokens arrive – no instruction locality  no use of registers

25. 25 Explicit Token Store (ETS) approach Target: efficient implementation of token matching. Basic idea: allocate a separate frame in a frame memory for each active loop iteration or subprogram invocation. A frame consists of slots; each slot holds an operand that is used in the corresponding activity. Access to slots is direct (i.e. through offsets relative to the frame pointer)  no associative search is needed.

26. 26 Explicit token store

27. 27 Monsoon, an explicit token store machine Processing Element Multistage Packet Switching Network PE I-Structure Storage I-Structure Storage... Frame Memory Form Token User QueueSystem Queue Instruction Memory ALU Instruction Fetch Effective Address Generation Presence Bit Operation Frame Operation to/from the Communication Network

28. 28 Monsoon, an explicit token store machine Each PE is using an eight-stage pipeline – instruction fetch --- precedes token matching (in contrast to dynamic dataflow processors with associative matching units)! – token matching1 - effective address generation: explicit token address is computed from the frame address and operand offset – token matching2 - presence bit operation: a presence bit is accessed to find out if the first operand of a dyadic operation has already arrived not arrived  presence bit is set and the current token is stored into the frame slot of the frame memory arrived  presence bit is reset and the operand can be retrieved from the slot of the frame memory in next stage – token matching3 - frame operation stage: Operand storing or retrieving. – Next three stages - execution stages in the course of which the next tag is also computed concurrently. – Eighth stage - form-token: forms one or two new tokens that are sent to the network, stored in a user token queue, a system token queue, or directly recirculated to the instruction fetch stage of the pipeline.

29. 29 Monsoon prototype 16 prototypes at beginning of 90ies! Processing element: – 10 MHz clock – 56 kW Instruction Memory (32 bit wide) – 256 kW Frame Memory (word + 3 presence bits, word size: 64 bit data + 8 bit tag) – Two 32 k token queues (system, user) I-structure storage: – 4MW (word + 3 presence bits) – 5 M requests/sec Network – Multistage, pipelined – Packet Routing Chips (PaRC, 4 x 4 crossbar) – 4 M tokens/s/link (100 MB/s)

30. 30 Dataflow processors - Hybrids Poor sequential code performance by dynamic dataflow computers – Why? – an instruction of the same thread is issued to the dataflow pipeline after the completion of its predecessor instruction. – In the case of an 8-stage pipeline, instructions of the same thread can be issued at most every eight cycles. – Low workload: the utilization of the dataflow processor drops to one eighth of its maximum performance. Another drawback: the overhead associated with token matching. – before a dyadic instruction is issued to the execution stage, two result tokens have to be present. – The first token is stored in the waiting-matching store, thereby introducing a bubble in the execution stage(s) of the dataflow processor pipeline. – measured pipeline bubbles on Monsoon: up to 28.75 % No use of registers possible!

31. 31 Augmenting dataflow with control-flow Solution: combine dataflow with control-flow mechanisms. – threaded dataflow, – large-grain dataflow, – dataflow with complex machine operations, – further hybrids.

32. 32 Threaded dataflow Threaded dataflow: the dataflow principle is modified so that instructions of certain instruction streams are processed in succeeding machine cycles. A subgraph that exhibits a low degree of parallelism is transformed into a sequential thread. The thread of instructions is issued consecutively by the matching unit without matching further tokens except for the first instruction of the thread. Threaded dataflow covers – the repeat-on-input technique used in Epsilon-1 and Epsilon-2 processors, – the strongly connected arc model of EM-4, and – the direct recycling of tokens in Monsoon.

33. 33 Threaded dataflow (continued) Data passed between instructions of the same thread is stored in registers instead of written back to memory. Registers may be referenced by any succeeding instruction in the thread. – Single-thread performance is improved. – The total number of tokens needed to schedule program instructions is reduced which in turn saves hardware resources. – Pipeline bubbles are avoided for dyadic instructions within a thread. Two threaded dataflow execution techniques can be distinguished: – direct token recycling (Monsoon), – consecutive execution of the instructions of a single thread (Epsilon & EM).

34. 34 Direct token recycling of Monsoon Cycle-by-cycle instruction interleaving of threads similar to multithreaded von Neumann computers! 8 register sets can be used by 8 different threads. Dyadic instructions within a thread (except for the start instruction!) refer to at least one register  need only a single token to be enabled. A result token of a particular thread is recycled ASAP in the 8-stage pipeline, i. e., every 8th cycle the next instruction of a thread is fired and executed. This implies that at least 8 threads must be active for a full pipeline utilization. Threads and fine-grain dataflow instructions can be mixed in the pipeline.

35. 35 Epsilon and EM-4 Instructions of a thread are executed consecutively. The circular pipeline of fine-grain dataflow is retained. The matching unit is enhanced with a mechanism that, after firing the first instruction of a thread, delays matching of further tokens in favor of consecutive issuing of all instructions of the started thread. Problem: implementation of an efficient synchronization mechanism

36. 36 Large-grain (coarse-grain) dataflow A dataflow graph is enhanced to contain fine-grain (pure) dataflow nodes and macro dataflow nodes. – A macro dataflow node contains a sequential block of instructions. A macro dataflow node is activated in the dataflow manner, its instruction sequence is executed in the von Neumann style! Off-the-shelf microprocessors can be used to support the execution stage. Large-grain dataflow machines typically decouple the matching stage (sometimes called signal stage, synchronization stage, etc.) from the execution stage by use of FIFO-buffers. Pipeline bubbles are avoided by the decoupling and FIFO-buffering.

37. 37 Dataflow with complex machine operations Use of complex machine instructions, e.g. vector instructions – ability to exploit parallelism at the subinstruction level – Instructions can be implemented by pipeline techniques as in vector computers. – The use of a complex machine operation may spare several nested loops. Structured data is referenced in block rather than element-wise and can be supplied in a burst mode.

38. 38 Dataflow with complex machine operations and combined with LGDF Often use of FIFO-buffers to decouple the firing stage and the execution stage – bridges different execution times within a mixed stream of simple and complex instructions. – Major difference to pure dataflow: tokens do not carry data (except for the values true or false). – Data is only moved and transformed within the execution stage. – Applied in: Decoupled Graph/Computation Architecture, the Stollmann Dataflow Machine, and the ASTOR architecture. – These architectures combine complex machine instructions with large-grain dataflow.

39. 39 Augmenting dataflow with control-flow

40. 40 Lessons learned from dataflow Superscalar microprocessors display an out-of-order dynamic execution that is referred to as local dataflow or micro dataflow. Colwell and Steck 1995, in the first paper on the PentiumPro: “The flow of the Intel Architecture instructions is predicted and these instructions are decoded into micro-operations (  ops), or series of  ops, and these  ops are register-renamed, placed into an out-of-order speculative pool of pending operations, executed in dataflow order (when operands are ready), and retired to permanent machine state in source program order.” State-of-the-art microprocessors typically provide 32 (MIPS R10000), 40 (Intel PentiumPro) or 56 (HP PA-8000) instruction slots in the instruction window or reorder buffer. Each instruction is ready to be executed as soon as all operands are available.

41. 41 Comparing dataflow computers with superscalar microprocessors Superscalar microprocessors are von Neumann based: (sequential) thread of instructions as input  not enough fine-grained parallelism to feed the multiple functional units  speculation Dataflow approach resolves any threads of control into separate instructions that are ready to execute as soon as all required operands become available. The fine-grained parallelism generated by dataflow principle is far larger than the parallelism available for microprocessors. However, locality is lost  no caching, no registers

42. 42 Lessons learned from dataflow (Pipeline issues) Microprocessors: Data and control dependences potentially cause pipeline hazards that are handled by complex forwarding logic. Dataflow: Due to the continuous context switches, pipeline hazards are avoided; disadvantage: poor single thread performance. Microprocessors: Antidependences and output dependences are removed by register renaming that maps the architectural registers to the physical registers. Thereby the microprocessor internally generates an instruction stream that satisfies the single assignment rule of dataflow. The main difference between the dependence graphs of dataflow and the code sequence in an instruction window of a microprocessor: branch prediction and speculative execution. Microprocessors: rerolling execution in case of a wrongly predicted path is costly in terms of processor cycles.

43. 43 Lessons learned from dataflow (Continued) Dataflow: The idea of branch prediction and speculative execution has never been evaluated in the dataflow environment. Dataflow was considered to produce an abundance of parallelism while speculation leads to speculative parallelism which is inferior to real parallelism. Microprocessors: Due to the single thread of control, a high degree of data and instruction locality is present in the machine code. Microprocessors: The locality allows to employ a storage hierarchy that stores the instructions and data potentially executed in the next cycles close to the executing processor. Dataflow: Due to the lack of locality in a dataflow graph, a storage hierarchy is difficult to apply.

44. 44 Lessons learned from dataflow (Continued) Microprocessors: The operand matching of executable instructions in the instruction window is restricted to a part of the instruction sequence. Because of the serial program order, the instructions in this window are likely to become executable soon.  The matching hardware can be restricted to a small number of slots. Dataflow: the number of tokens waiting for a match can be very high.  A large waiting-matching store is required. Dataflow: Due to the lack of locality, the likelihood of the arrival of a matching token is difficult to estimate,  caching of tokens to be matched soon is difficult.

45. 45 Lessons learned from dataflow (Memory latency) Microprocessors: An unsolved problem is the memory latency caused by cache misses. Example: SGI Origin 2000: – latencies are 11 processor cycles for a L1 cache miss, – 60 cycles for a L2 cache miss, – and can be up to 180 cycles for a remote memory access. – In principle, latencies should be multiplied by the degree of superscalar. Microprocessors: Only a small part of the memory latency can be hidden by out-of-order execution, write buffer, cache preload hardware, lockup free caches, and a pipelined system bus. Microprocessors often idle and are unable to exploit the high degree of internal parallelism provided by a wide superscalar approach. Dataflow: The rapid context switching avoids idling by switching execution to another context.

46. 46 Lessons learned from dataflow (Continued) Microprocessors: Finding enough fine-grain parallelism to fully exploit the processor will be the main problem for future superscalars. Solution: enlarge the instruction window to several hundred instruction slots; two draw-backs – Most of the instructions in the window will be speculatively assigned with a very deep speculation level (today's depth is normally four at maximum).  most of the instruction execution will be speculative. The principal problem here arises from the single instruction stream that feeds the instruction window. – If the instruction window is enlarged, the updating of the instruction states in the slots and matching of executable instructions lead to more complex hardware logic in the issue stage of the pipeline thus limiting the cycle rate.

47. 47 Lessons learned from dataflow (Continued) Solutions: – the decoupling of the instruction window with respect to different instruction classes, – the partitioning of the issue stage into several pipeline stages, – and alternative instruction window organizations. Alternative instruction window organization: the dependence-based microprocessor: – Instruction window is organized as multiple FIFOs. – Only the instructions at the heads of a number of FIFO buffers can be issued to the execution units in the next cycle. – The total parallelism in the instruction window is restricted in favor of a less costly issue that does not slow down processor cycle rate. – Thereby the potential fine-grained parallelism is limited  somewhat similar to the threaded dataflow approach.

48. 48 Lessons learned from dataflow (alternative instruction window organizations) Look at dataflow matching store implementations Look into dataflow solutions like threaded dataflow (e.g. repeat-on-input technique or strongly-connected arcs model) Repeat-on-input strategy issues compiler-generated code sequences serially (in an otherwise fine-grained dataflow computer).  Transferred to the local dataflow in an instruction window: – an issue string might be used; – a serie of data dependent instructions is generated by a compiler and issued serially after the issue of the leading instruction. However, the high number of speculative instructions in the instruction window remains.