Stefan J. Geuns

Dataflow analysis for multiprocessor systems with non-starvation-free schedulers

Dataflow analysis techniques are suitable for the temporal analysis of real-time stream processing applications. However, the applicability of these models is currently limited to systems with starvation-free schedulers, such as Time-Division Multiplexing (TDM) schedulers. Removal of this limitation would broaden the application domain of dataflow analysis techniques significantly. In this paper we present a temporal analysis technique for Homogeneous Synchronous Dataflow (HSDF) graphs, that is also applicable for systems with non-starvation-free schedulers. Unlike existing dataflow analysis techniques, the proposed analysis technique makes use of an enabling-jitter characterization and iterative fixed-point computation. The presented approach is applicable for arbitrary (cyclic) graph topologies. Buffer capacity constraints are taken into account during the analysis and sufficient buffer capacities can be determined afterwards. The approach presented in this paper is the first approach that considers non-starvation-free schedulers in combination with arbitrary HSDF graphs The proposed dataflow analysis technique is implemented in a tool. This tool is used to evaluate the analysis technique using examples that illustrate some important differences with other temporal analysis methods. The case-study discusses how the method presented in this paper can be used to solve a problem with the inaccuracy of the temporal analysis results of a real-time stream processing system. This stream processing system consists of an FM receiver together with a DAB receiver application which both share a Digital Signal Processor (DSP).

Compositional temporal analysis model for incremental hard real-time system design

The incremental design and analysis of parallel hard real-time stream processing applications is hampered by the lack of an intuitive compositional temporal analysis model that supports arbitrary cyclic dependencies between tasks. This paper introduces a temporal analysis model for hard real-time systems, called the Compositional Temporal Analysis (CTA) model, in which arbitrary cyclic dependencies can be specified. The CTA model also supports hierarchical composition and incremental design of timed components. The internals of a component in the CTA model can be hidden without changing the temporal properties of the component. Furthermore, the composition operation in the CTA model is associative, which enables composing components in an arbitrary order. Besides all these properties, also latency constraints and periodic sources and sinks can be specified and analyzed. We also show in this paper that for the CTA model efficient algorithms exist for buffer sizing, verifying consistency of compositions and to compute the temporal properties of compositions. The CTA model can be used as an abstraction of timed dataflow models. The CTA model uses components with transfer rates per port, in contrast to dataflow models that use actors with firing rules. Unlike dataflow models, the CTA model is not executable. An audio echo cancellation application is used to illustrate the applicability of the CTA model for a stream processing application with throughput and latency constraints, and to illustrate incremental design.

Automatic dataflow model extraction from modal real-time stream processing applications

Many real-time stream processing applications are initially described as a sequential application containing while-loops, which execute for an unknown number of iterations. These modal applications have to be executed in parallel on an MPSoC system in order to meet their real-time throughput constraints. However, no suitable approach exists that can automatically derive a temporal analysis model from a sequential specification containing while- loops with an unknown number of iterations. This paper introduces an approach to the automatic generation of a Structured Variable-rate Phased Dataflow (SVPDF) model from a sequential specification of a modal application. The real-time requirements of an application can be analyzed despite the presence of while-loops with an unknown number of iterations. It is shown that an algorithm that has a polynomial time computational complexity can be applied on the generated SVPDF model to determine whether a throughput constraint can be met. The enabler for the automatic generation of an SVPDF model is the decoupling of synchronization between tasks that contain different while-loops. A DVB-T radio transceiver illustrates the derivation of the SVPDF model.

Buffer Sizing to Reduce Interference and Increase Throughput of Real-Time Stream Processing Applications

Existing temporal analysis and buffer sizing techniques for real-time stream processing applications ignore that FIFO buffers bound interference between tasks on the same processor. By considering this effect it can be shown that a reduction of buffer capacities can result in a higher throughput. However, the relation between buffer capacities and throughput is non-monotone in general, which makes an exploitation of the effect challenging. In this paper a buffer sizing approach is presented which exploits that FIFO buffers bound interference between tasks on shared processors. The approach combines temporal analysis using a cyclic dataflow model with computation of buffer capacities in an iterative manner and thereby enables higher throughput guarantees at smaller buffer capacities. It is shown that convergence of the proposed analysis flow is guaranteed. The benefits of the presented approach are demonstrated using a WLAN 802.11p transceiver application executed on a multiprocessor system with shared processors. If buffers without blocking writes are used an up to 25% higher guaranteeable throughput and up to 23% smaller buffer capacities can be determined compared to existing approaches. For systems using buffers with blocking writes the guaranteeable throughput is even up to 43% higher and buffer capacities up to 11% smaller.

Temporal analysis flow based on an enabling rate characterization for multi-rate applications executed on mpsocs with non-starvation-free schedulers

Real-time stream processing applications often contain multi-rate behavior. This multi-rate behavior can be accurately modeled using Synchronous Dataflow (SDF) graphs. However, no temporal analysis technique exists which is applicable for arbitrary cyclic SDF graphs and can handle cyclic resource dependencies. This paper presents a temporal analysis flow for SDF graphs which is applicable for systems with non-starvation-free schedulers such as static priority pre-emptive schedulers. The analysis flow uses an enabling rate characterization to calculate response times. This enabling rate characterization is determined using multi-dimensional periodic schedules and allows a more accurate modeling of enabling patterns than is possible with a characterization that is based on periods and enabling jitters. The presented approach is applicable for arbitrary (cyclic) graph topologies and can take buffer capacity constraints into account during analysis. Also cyclic resource dependencies can be analyzed. The presented analysis flow is the first approach that considers arbitrary SDF graph topologies in combination with cyclic resource dependencies that are caused by non-starvation-free schedulers. The proposed analysis flow is evaluated using a radio processing application. The analysis results are obtained using a tool in which the analysis flow is implemented. This case-study illustrates that the used enabling characterization achieves up to 87% better response times than with an enabling jitter based characterization.

Two parameter workload characterization for improved dataflow analysis accuracy

Real-time stream processing applications, such as radios, can often be modeled intuitively with dataflow models. Given the Worst-Case Execution Times (WCETs) of the tasks, which characterizes workload with one parameter, dataflow analysis techniques have been used to compute the minimum throughput and maximum latency of these applications. However, a large difference between the WCETs of the tasks and their average execution times can result in a large difference between the computed worst-case throughput and the actual obtained throughput. To reduce the difference between the worst-case throughput, determined by analysis, and the actual obtained throughput, we introduce in this paper a two parameter (σ,ρ) workload characterization of the tasks to improve the accuracy of dataflow analysis. The (σ, ρ) workload characterization captures information on the maximum cumulative execution time of consecutive executions of a task and can therefore be seen as a generalization of the WCET characterization. We show how the (σ,ρ) workload characterization can be used in combination with several types of dataflow graphs and how it can be used to improve the temporal analysis results of real-time stream processing applications. We illustrate this for a DVB-T radio application, a car-radio application and a data-dependent MP3 playback application.

Accuracy Improvement of Dataflow Analysis for Cyclic Stream Processing Applications Scheduled by Static Priority Preemptive Schedulers

Stream processing applications executed on embedded multiprocessor systems regularly contain cyclic data dependencies due to the presence of feedback loops and bounded FIFO buffers. Dataflow modeling is suitable for the temporal analysis of such applications. However, the accuracy can be unsatisfactory as existing temporal analysis techniques ignore that cyclic data dependencies limit interference between tasks executed on shared processors. This paper presents a dataflow analysis approach that increases the analysis accuracy by taking into account that cyclic data dependencies limit interference between tasks. It is shown that the approach is applicable for single-rate stream processing applications that are executed on multiprocessor systems using static priority preemptive schedulers. The improvement of accuracy is demonstrated in a case study employing a WLAN 802.11p transceiver application that is executed on a multiprocessor system with shared processors.

Temporal analysis model extraction for optimizing modal multi-rate stream processing applications

Modern real-time stream processing applications, such as Software Defined Radio (SDR) applications, typically have multiple modes and multi-rate behavior. Modes are often described using while-loops whereas multi-rate behavior is frequently described using arrays with pseudo-random indexing patterns. The temporal properties of these applications have to be analyzed in order to determine whether optimizations improve throughput. However, no method exists in which a temporal analysis model is derived from these applications that is suitable for temporal analysis and optimization. In this paper an approach is presented in which a concurrency model for the temporal analysis and optimization of stream processing applications is automatically extracted from a parallelized sequential application. With this model it can be determined whether a program transformation improves the worst-case temporal behavior. The key feature of the presented approach is that arrays with arbitrary indexing patterns can be described, allowing the description of multi-rate behavior, while still supporting the description of modes using while-loops. In the model, an over-approximation of the synchronization dependencies is used in case of arrays with pseudo-random indexing patterns. Despite the use of this approximation, we show that deadlock is only concluded from the model if there is also deadlock in the parallelized application. The relevance and applicability of the presented approach are demonstrated using an Orthogonal Frequency-Division Multiplexing (OFDM) transmitter application.

Unified dataflow model for the analysis of data and pipeline parallelism, and buffer sizing

Real-time stream processing applications such as software defined radios are usually executed concurrently on multiprocessor systems. Exploiting coarse-grained data parallelism by duplicating tasks is often required, besides pipeline parallelism, to meet the temporal constraints of the applications. However, no unified model and analysis method exists that can be used to determine the required amount of data and pipeline parallelism, and buffer sizes simultaneously. This paper presents an analysis method which can determine the required amount of data parallelism by describing data parallelism in a dataflow model without replicating dataflow actors. This allows to make a trade-off between the amount of data and pipeline parallelism that is required to meet the temporal constraints of the application. It is also shown how large the buffers need to be such that the determined amount of data and pipeline parallelism required for the satisfaction of the throughput constraint, can be realized. Furthermore, it is shown that the use of the applied circular buffers enables the proposed dataflow modeling. The presented analysis method is demonstrated using a WLAN 802.11p transceiver application. This application contains multi-rate behavior and has a cyclic data dependency because of a re-encoding loop. Given the real-time constraints of the application, sufficient buffer sizes and sufficient data parallelism are derived.

Sequential specification of time-aware stream processing applications

Automatic parallelization of Nested Loop Programs (NLPs) is an attractive method to create embedded real-time stream processing applications for multi-core systems. However, the description and parallelization of applications with a time dependent functional behavior has not been considered in NLPs. In such a description, semantic information about time dependent behavior must be made available for the compiler, such that an optimized time independent implementation can be generated automatically. This article introduces language constructs with temporal semantics to NLPs. Using these language constructs, time dependent applications can be specified and a corresponding data-driven implementation can be generated for use on a multi-core system. Despite that these time-aware language constructs can be data-dependent, the application remains functionally deterministic. Pipelining is exploited to increase the throughput of an application. The media access control (MAC) protocol of an IEEE 802.11p WLAN transceiver is used to illustrate the relevance and applicability of the introduced concepts.