Programming many core systems Marco Bekooij Outline Definition many core systems Application domain of many core systems Microsoft Parallel Computing Initiative – simplify programming – improve quality of service Mapping stream processing on real-time multiprocessor systems – Automatic parallelization – Budget computation – Multiprocessor system hardware design with budget enforcement Conclusion Definition many core system according to Intel’s white paper Many core systems are multiprocessor systems with a large number of cores (>8) – Many core systems have a shared address space and its resources are under control of the operating system Computing industry shifts and their effects on user experience Parallel computing and the next generation of user experiences Microsoft: many core applications Next-Generation Personal Computing Experiences – Personal modeling: e.g. “walk-through” 3-dimensional, photo-realistic renderings of a home renovation – Personalized adaptive learning: create a personalized, context-aware curriculum in real-time. – Public safety: detailed 2- or 3-dimensional renderings, object recognition, help responders to make well-informed decisions critical to rescue tactics, evacuations, and emergency response Business Opportunities – Financial modeling – Product design simulation Many core application example of NXP Multi-stream multi-standard car-infotainment systems – Advanced radios contain already 13 processors (10 DSPs + 3 µP) + number of hardware accelerators Beamforming Improved radio reception Microsoft Parallel Computing Initiative Objectives: • simplify parallel software development • take quality-of-service requirements into account Microsoft Parallel Computing Initiative Applications: next experiences, improve productivity Domain libraries: system building blocks for example imageprocessing libraries Programming models and languages: easy application development without the need for expert knowledge Developer tooling: simplify software integration Runtime, platform, and operating systems: more effectively budget and arbitrate among competing requests for available resources in the face of parallelism and quality-of-service demands. Additionally, Microsoft will continue to improve the reliability and security of the platform. NXP’s application domain: real-time stream processing use-case f2 f1 ADC PDC CFE VIT CBE SRC APP digital radio job A f3 BR use-case MP3 SRC source decoding job B APP DAC DAC Software mapping flow for real-time stream processing applications Temporal constraints Architecture instance NLP Omphale (parallelization) Task graph +Dataflow graph Execution time analysis Task-graph + Dataflow graph Hebe (budget computation) Task-graph + budgets Use-cases + transitions Off-line = at design-time On-line = at run-time Helios (resource allocation) Table with resource allocations Start/stop job Minos (resource assignment) Preemptive kernel FIFO com lib Run-time mapping of tasks to processors with admission control per job Temporal constraints Architecture instance NLP Omphale (parallelization) Task graph +Dataflow graph Execution time analysis Task-graph + Dataflow graph Hebe (budget computation) Task-graph + budgets Off-line = at design-time On-line = at run-time Start/stop job Minos (resource assignment) Preemptive kernel FIFO com lib Setting computation requires property preserving abstraction Budget computation Dataflow model v1 v0 v2 Abstraction DSP P $ External SDRAM I/O ctrl mem NI NI Network NI NI Experimental predictable many core system Timer Blaze ROM MEM Timer Blaze ROM $ MEM RS232 $ Aethereal NoC SDRAM ctrl Distributed shared memory system – Pthread support SDRAM Budget scheduler for every shared resource – processors, memory ports, inter-connect Flow control – back-pressure Mapped on a Vitex 4 FPGA Heterogeneous many core system Timer DSP ROM IMEM Timer Blaze ROM DMEM MEM RS232 $ Aethereal NoC Heterogeneous for area-efficiency and power-efficiency reasons Streaming without addresses over the network beside address based streaming SDRAM ctrl SDRAM Essential elements in the approach Key assumption: characteristics of other jobs are not completely known at design time: – Other jobs are downloaded – Worst-case execution times of the tasks are not known at design time Essential element – Budget schedulers – Flow control Budget schedulers All schedulers Budget schedulers Budget scheduler: subclass of the aperiodic server – minimum budget in a replenishment interval is independent of the execution-time and event arrival-rate Budget reservation: – incomplete knowledge: worst-case execution times of the tasks of other jobs are not known – overload protection: estimated execution times are optimistic Budget scheduler example: time division multiplex x(j): execution time of the j-th execution, P: period, B: slice length Budget scheduler with priorities: PBS • number of preemptions in a RI is fixed preemption overhead is known • maximum time between event and start of high priority task with budget Flow control Data can be lost without flow control non deterministic functional behavior Buffer overflow P Buffer overflow can occur if: C – best-case execution time of producer P is over estimated – worst-case execution time of consumer C is under estimated Task graph and dataflow graph extration Extraction of a task graph is difficult – Data dependency analysis Derivation of an (dataflow) analysis model of an application is difficult an error prone – No one-to-one correspondence between task graph and dataflow graph Our approach: describe top-level of the application as a nested loop program – Allow while loops, if conditions, and non-affine index expressions Nested loop programs (NLPs) A nested loop program is specified in a coordination language – Specifies dependencies (communication) between functions – functions are defined in a programming language, e.g. C – to simplify/enable parallelization • many programming language constructs are not supported to simplify/enable analysis • new program language constructs have been added to improve the analyzability (and therefore NLPs are not an C-subset) Nested loop programs should be seen as a sequential specification of a task graph – single assignment • each array location is written at most once during one execution of the outer while loop – functions must be side-effect free Nested loop program example mode=0; while(1){ in=input(); switch(mode){ case 0: {mode=detect(in); } case 1: {mode,o1=decode1(in); o2=decode2(o1); output(o2);} } } Resulting task graph det in dec1 dec2 out Every function becomes a task Buffers can have multiple readers Buffers can have multiple mutual exclusive writers – That writes are mutual exclusive is explicit in the NLP but not in the task-graph Budget computation Budgets are computed given real-time constraints – only end-2-end constraints are imposed by the environment • throughput + latency and not the deadlines of the tasks Requires an suitable analysis model for real-time applications – should take pipelining into account • the i-th input sample is consumed before the (i-1)th output sample is produced We apply dataflow analysis with measured (not worst-case) execution times Definition real-time system Real-time systems are those systems in which the correctness of the system depends not only on the logical results of computation, but also on the time at which the results are produced. Real-time analysis Use of measured execution times instead of worst-case execution times – Guarantees? – Load hypothesis Basics of dataflow analysis Distinguishing features of real-time systems High level of determinism: – It should be possible to derive useful properties of the system, given the stated assumptions and the information available with an acceptable effort and a useful accuracy Concurrency: – Deal with the inherent physical concurrency – Deal with a concurrent description of the system – Deal with a concurrent implementation of the system Emphasis and significance of reliability and fault tolerance: – Reliability is the probability that a system will perform correctly over a given period of time – Fault tolerance is concerned with the recognition and handling of failures Computer Assisted Control Definition predictability Is should be possible to show, demonstrate, or prove that requirements are met subject to assumptions made, for example, concerning failures and workloads. Note that: – predictability is always subject to the underlying assumptions made Real-time system classification Note that: • no deadlines are defined for best-effort tasks • assumes that all tasks in the system have the same criticality Criticality spectrum for systems Hard RT Very critical Firm RT Soft RT Best effort Not critical at all Load hypothesis Statement about the assumption of the peak load of the system Translates often in an assumption about the worst-case execution times of the tasks Difference between guarantee and a statistical assertion A guarantee is an assurance of a fulfillment of a condition – a guarantee is binary statement – guarantees about the reality are given under certain assumptions Statistical assertion is a statement about a probability of an occurrence Focus is on analysis techniques that result in guarantees – guarantees are given under explicit and testable assumptions (in our case the load hypothesis) Research Schools 1. Real-time system theory should help to give guarantees about the temporal behavior of the system Testing can provide only a partial verification of the behavior. This justifies the use of analytical techniques that can provide complete coverage. Classical view of the real-time community 2. Real-time system theory should provide means to manage the system resources such that the temporal behavior improves 3. Real-time system theory should provide means to compute system settings 4. Real-time system theory should provide means to reduce the verification effort 5. Real-time system theory should provide means to improve the robustness of a system Load hypothesis for firm real-time systems Often execution times are measured instead of computed with WCET tools – reason: WCET tools are not available or computed WCETs are overly pessimistic Typically a load hypothesis is defined which states that the execution times of the tasks are not larger than the WCETs used during analysis Given that the load hypothesis holds we can guarantee with analysis techniques that no deadlines are missed If the load hypothesis does not hold then no statements can be made about the worst-case temporal behavior of the system Assumption coverage Strength of materials theory – Model is for example an approximation of a bridge • E.g. the stiffness of the metal beam is intrinsically not exactly known, i.e. can be worse or better – However model can be a useful approximation of the reality • Added safety margin (head room) is based on experience Does the same reasoning apply to real-time system design? False useful results? Usefulness real-time analysis results even given unsafe execution time estimates Deadlock freedom and functional determinism of the application Estimates of the real-time behavior of the tasks Estimates of appropriate system configuration and system settings Trends and anomalies Responsiveness improvement Sensitivity reduction Robustness improvement Synchronization and scheduling overhead reduction Focus of real-time analysis techniques Single processor – Focus is on task scheduling of independent tasks + OS-kernel design Multiprocessor – Focus is on throughput and latency analysis of applications described as task graphs • also synthesis of settings & budget such that throughput and latency constraints are met Formal models for real-time analysis Process algebra – Algebra for communicating processes – Allows transformation of one system into another Temporal logics – Propositional logic augmented by tense operators – System representation with global states become prohibitive large Automata – Mathematical model for a finite statemachines – Synchrony timing hypothesis OR clocks • instantaneous broad-cast • system evolves faster than events Petri nets – Dataflow graphs have similarities with Petri nets Classical timing verification techniques Logic based approach – Deductive proof: IS or decision procedure Inot(S) is unsatisfiable – Very high computational complexity and hard to automate Automata based approach – Language containment: Li Ls – State explosion Model checking – State explosion – High computational complexity Outside scope of these formal verification techniques Techniques to include resource sharing – effects of scheduling on the temporal behavior Techniques to make an abstraction of the system while preserving properties Techniques to synthesize properties instead of checking properties Techniques to trade accuracy for lower computational complexity Techniques to trade expressivity model for analyzability Techniques to trade generality model for analyzability However these techniques are essential for real-time multiprocessor system design – borrow ideas from performance analysis of communication networks – Latency rate-analysis dataflow analysis Rate based analysis Rate based analysis determines, loosely speaking, the throughput of the system Three approaches: 1. Graph-based techniques: maximum cycle mean analysis 2. Algebraic techniques: determine eigenvectors with max-plus algebra – Stochastic approaches: Markov process Limitations: – – – 1 and 2 assumes fixed timing delays instead of intervals, while 3 computes the for real-time systems not very useful long term average throughput supported models do not support any choice supported models do not support inputs and outputs Still useful for real-time analysis purposes? (yes) Are there solutions available to analyze data dependent applications? (yes) Related work: worst-case performance analysis of communication networks Objective compute maximum latency and minimum throughput for a flow of packets Links between routers are shared by flows No flow control: input buffers must be large enough such that overflow does not occur Related work: worst-case performance analysis [R. Cruz, 1991]: A Calculus for Network Delay – No flow control, does not require starvation free schedulers – Bound traffic for t0 with a non-decreasing function [K. Tindell and J. Clark, 1994]: Holistic Schedulability Analysis – No flow control, static priority preemptive – fixed point iteration in case of cyclic resource dependencies [D. Stiliadis et.al., 1998]: Latency-Rate Servers: A General Model for Analysis of Traffic Scheduling Algorithms – No flow control – Requires starvation-free schedulers there are no cyclic resource dependencies – System can be characterize without knowledge about the input traffic • use of the concept of busy periods – More accurate estimate of the end-to-end delay than [Cruz91] and [TC94] [J.Y. Le Boudec, 1998]: Application of Network Calculus to Guaranteed Service Networks – No flow control – Does not require schedulers to be starvation-free – More accurate estimate of the end-to-end delay than [Cruz91] and [TC94] Related work: worst-case performance analysis of task graphs [RZJE02, JRE02] Event model composition – Definition of period+jitter traffic models, tasks with AND-condition, generalization of [TC94] [S. Chakraborty et.al., 2003] A General framework for .... – Generalization network calculus, also known as real-time calculus – Bound traffic and service for any interval t – Acyclic task graphs [M.Wiggers, et.al., 2007] Modeling Run-Time Arbitration by Latency-Rate Servers in Data Flow Graphs – Requires starvation-free schedulers – Applicable in case of arbitrary deterministic task graphs: AND, cyclic task graphs, buffer capacity can be given or computed [M.Wiggers, et.al., 2009] Monotonicity and run-time scheduling – Generalization of [M.Wiggers, and M. Bekooij, 2007]: allows sequence of execution times and any deterministic dataflow graph – not based on busy periods [L. Thiele, and M. Stoimenov, 2009] Modular performance analysis of cyclic dataflow graphs – Generalization real-time calculus [S. Chakraborty et.al., 2003] : Analysis of cyclic HSDF graphs Related work: worst-case performance analysis of task graphs [J. Staschulat, et.al. 2009] Dataflow models for shared memory access latency analysis – piece-wise linear service approximation of priority based budget schedulers [M. Wiggers, et.al., 2010] Simultaneous Budget and Buffer Size Computation for Throughput-Constrained Task Graphs – only HSDF graphs – algorithm has a polynomial computational complexity Dataflow analysis primer Elements in the single-rate dataflow model Actors An actor is depicted as a node An actor is stateless An actor can represent a function An actor can be use to represent a task (but also other things) Actor fire (a task execute) Actors have a firing conditions A firing duration can be associated with an actor Actors only interact with their environment through token consumption from their input queues and token production through their output queues Queues A queue is represented by an edge Queues have per definition an unbounded capacity Tokens are stored in a queue Tokens can be consumed from a queue in the order that they are produced Tokens A token is an undividable element Tokens can be use to represent: data, space, or synchronization moments Firing rule A firing rule is a condition that prescribes the number of tokens that must be present in the input queues of an actor before the actor can fire. The firing rule of a single rate dataflow actor (also called HSDF actor) is: – one token in each input queue Notice: a firing rule cannot specify anything about the number of tokens in an output queue Throughput calculation example What is the throughput during self-timed execution? Token arrival times during self-timed executed HSDF Given an HSDF graph G(V,E) The self-timed execution of this graph has some important properties if: – The graph is strongly connected, i.e. there is a directed path from every node to every other node in the graph – Actors have a constant firing duration It can be shown that the graph enters a periodic regime after an initial transition phase [BCOQ92] Multi-dimensional periodic schedule On average every a firing Maximum cycle mean [Rei68] Maximum cycle mean example MCM number example (1) The critical cycle determines the mcm The nodes and edges colored red belong to the critical cycle MCM number example (2) Overlapping firings MCM calculation example What is the mcm of this graph? Related work: worst-case performance analysis of communication networks Objective compute maximum latency and minimum throughput for a flow of packets Links between routers are shared by flows No flow control: input buffers must be large enough such that overflow does not occur Flow control Related work: worst-case performance analysis [R. Cruz, 1991]: A Calculus for Network Delay – No flow control, does not require starvation free schedulers – Bound traffic for t0 with a non-decreasing function [K. Tindell and J. Clark, 1994]: Holistic Schedulability Analysis – No flow control, static priority preemptive – fixed point iteration in case of cyclic resource dependencies [D. Stiliadis et.al., 1998]: Latency-Rate Servers: A General Model for Analysis of Traffic Scheduling Algorithms – No flow control – Requires starvation-free schedulers there are no cyclic resource dependencies – System can be characterize without knowledge about the input traffic • use of the concept of busy periods – More accurate estimate of the end-to-end delay than [Cruz91] and [TC94] [J.Y. Le Boudec, 1998]: Application of Network Calculus to Guaranteed Service Networks – No flow control – Does not require schedulers to be starvation-free – More accurate estimate of the end-to-end delay than [Cruz91] and [TC94] Related work: worst-case performance analysis of task graphs [RZJE02, JRE02] Event model composition – Definition of period+jitter traffic models, tasks with AND-condition, generalization of [TC94] [S. Chakraborty et.al., 2003] A General framework for .... – Generalization network calculus, also known as real-time calculus – Bound traffic and service for any interval t – Acyclic task graphs [M.Wiggers, et.al., 2007] Modeling Run-Time Arbitration by Latency-Rate Servers in Data Flow Graphs – Requires starvation-free schedulers – Applicable in case of arbitrary deterministic task graphs: AND, cyclic task graphs, buffer capacity can be given or computed [M.Wiggers, et.al., 2009] Monotonicity and run-time scheduling – Generalization of [M.Wiggers, and M. Bekooij, 2007]: allows sequence of execution times and any deterministic dataflow graph – not based on busy periods [L. Thiele, and M. Stoimenov, 2009] Modular performance analysis of cyclic dataflow graphs – Generalization real-time calculus [S. Chakraborty et.al., 2003] : Analysis of cyclic HSDF graphs Related work: worst-case performance analysis of task graphs [J. Staschulat, et.al. 2009] Dataflow models for shared memory access latency analysis – piece-wise linear service approximation of priority based budget schedulers [M. Wiggers, et.al., 2010] Simultaneous Budget and Buffer Size Computation for Throughput-Constrained Task Graphs – only HSDF graphs – algorithm has a polynomial computational complexity Classical dataflow models Fundamental differences dataflow models HSDF: Homogenous synchronous dataflow model – single-rate, polynomial MCM-algorithm SDF: Synchronous dataflow model – multi-rate, graph consistency, no known polynomial MCM-algorithm CSDF: cyclo-static dataflow model – fixed number of phases FDDF: functional deterministic dataflow model – data-dependent sequential firing rules, Turing complete (halting/deadlock) DDF: dynamic dataflow model – non-deteriministic firing rules Dataflow model subclasses DFM=data flow model subclass If DFMA DFMB then DFGA DFMA DFGA DFMB Recently introduced dataflow models with data dependent quanta Variable rate dataflow (VRDF) and variable rate phased dataflow (VPDF) is Turing complete how can we deal with undecidability? Hasse diagram representation including some recently introduced dataflow models Da Db if for every dataflow graph da Da it holds that da Db Conclusion Many core systems are – shared address space multiprocessor systems – number of core >8 A large number of cores put more stress on: – system programming effort – quality of service aspect Ongoing research effort on: – parallelization of a sequential description of stream processing algorithms – resource management: • computation of budgets • enforcement of budgets Computation of budgets – focus is on dataflow analysis techniques, challenges include • data dependent behavior (+ adaptation of resource budgets) • tradeoff between run-time and computational complexity