On a multi-radio baseband system, multiple independent transceivers must share the resources of a multi-processor, while meeting each its own hard real-time requirements. Not all possible combinations of transceivers are known at compile time, so a solution must be found that either allows for independent timing analysis or relies on runtime timing analysis. This thesis proposes a design flow and software architecture that meets these challenges, while enabling features such as independent transceiver compilation and dynamic loading, and taking into account other challenges such as ease of programming, efficiency, and ease of validation. We take data flow as the basic model of computation, as it fits the application domain, and several static variants (such as Single-Rate, Multi-Rate and Cyclo-Static) have been shown to possess strong analytical properties. Traditional temporal analysis of data flow can provide minimum throughput guarantees for a self-timed implementation of data flow. Since transceivers may need to guarantee strictly periodic execution and meet latency requirements, we extend the analysis techniques to show that we can enforce strict periodicity for an actor in the graph; we also provide maximum latency analysis techniques for periodic, sporadic and bursty sources. We propose a scheduling strategy and an automatic scheduling flow that enable the simultaneous execution of multiple transceivers with hard-realtime requirements, described as Single-Rate Data Flow (SRDF) graphs. Each transceiver has its own execution rate and starts and stops independently from other transceivers, at times unknown at compile time, on a multiprocessor. We show how to combine scheduling and mapping decisions with the input application data flow graph to generate a worst-case temporal analysis graph. We propose algorithms to find a mapping per transceiver in the form of clusters of statically-ordered actors, and a budget for either a Time Division Multiplex (TDM) or Non-Preemptive Non-Blocking Round Robin (NPNBRR) scheduler per cluster per transceiver. The budget is computed such that if the platform can provide it, then the desired minimum throughput and maximum latency of the transceiver are guaranteed, while minimizing the required processing resources. We illustrate the use of these techniques to map a combination of WLAN and TDS-CDMA receivers onto a prototype Software-Defined Radio platform. The functionality of transceivers for standards with very dynamic behavior – such as WLAN – cannot be conveniently modeled as an SRDF graph, since SRDF is not capable of expressing variations of actor firing rules depending on the values of input data. Because of this, we propose a restricted, customized data flow model of computation, Mode-Controlled Data Flow (MCDF), that can capture the data-value dependent behavior of a transceiver, while allowing rigorous temporal analysis, and tight resource budgeting. We develop a number of analysis techniques to characterize the temporal behavior of MCDF graphs, in terms of maximum latencies and throughput. We also provide an extension to MCDF of our scheduling strategy for SRDF. The capabilities of MCDF are then illustrated with a WLAN 802.11a receiver model. Having computed budgets for each transceiver, we propose a way to use these budgets for run-time resource mapping and admissibility analysis. During run-time, at transceiver start time, the budget for each cluster of statically-ordered actors is allocated by a resource manager to platform resources. The resource manager enforces strict admission control, to restrict transceivers from interfering with each other’s worst-case temporal behaviors. We propose algorithms adapted from Vector Bin-Packing to enable the mapping at start time of transceivers to the multi-processor architecture, considering also the case where the processors are connected by a network on chip with resource reservation guarantees, in which case we also find routing and resource allocation on the network-on-chip. In our experiments, our resource allocation algorithms can keep 95% of the system resources occupied, while suffering from an allocation failure rate of less than 5%. An implementation of the framework was carried out on a prototype board. We present performance and memory utilization figures for this implementation, as they provide insights into the costs of adopting our approach. It turns out that the scheduling and synchronization overhead for an unoptimized implementation with no hardware support for synchronization of the framework is 16.3% of the cycle budget for a WLAN receiver on an EVP processor at 320 MHz. However, this overhead is less than 1% for mobile standards such as TDS-CDMA or LTE, which have lower rates, and thus larger cycle budgets. Considering that clock speeds will increase and that the synchronization primitives can be optimized to exploit the addressing modes available in the EVP, these results are very promising.
|Kwalificatie||Doctor in de Filosofie|
|Datum van toekenning||10 jan 2012|
|Plaats van publicatie||Eindhoven|
|Status||Gepubliceerd - 2012|