Memory Hierarchy Hardware-Software Co-design in Embedded Systems

The memory hierarchy is the main bottleneck in modern computer systems as the gap between the speed of the processor and the memory continues to grow larger. The situation in embedded systems is even worse. The memory hierarchy consumes a large amount of chip area and energy, which are precious resources in embedded systems. Moreover, embedded systems have multiple design objectives such as performance, energy consumption, and area, etc. Customizing the memory hierarchy for specific applications is a very important way to take full advantage of limited resources to maximize the performance. However, the traditional custom memory hierarchy design methodologies are phase-ordered. They separate the application optimization from the memory hierarchy architecture design, which tend to result in local-optimal solutions. In traditional Hardware-Software co-design methodologies, much of the work has focused on utilizing reconfigurable logic to partition the computation. However, utilizing reconfigurable logic to perform the memory hierarchy design is seldom addressed. In this paper, we propose a new framework for designing memory hierarchy for embedded systems. The framework will take advantage of the flexible reconfigurable logic to customize the memory hierarchy for specific applications. It combines the application optimization and memory hierarchy design together to obtain a global-optimal solution. Using the framework, we performed a case study to design a new software-controlled instruction memory that showed promising potential.

Memory hierarchy and data communication in heterogeneous reconfigurable SoCs

The miniaturization race in the hardware industry aiming at continuous increasing of transistor density on a die does not bring respective application performance improvements any more. One of the most promising alternatives is to exploit a heterogeneous nature of common applications in hardware. Supported by reconfigurable computation, which has already proved its efficiency in accelerating data intensive applications, this concept promises a breakthrough in contemporary technology development. Memory organization in such heterogeneous reconfigurable architectures becomes very critical. Two primary aspects introduce a sophisticated trade-off. On the one hand, a memory subsystem should provide well organized distributed data structure and guarantee the required data bandwidth. On the other hand, it should hide the heterogeneous hardware structure from the end-user, in order to support feasible high-level programmability of the system. This thesis work explores the heterogeneous reconfigurable hardware architectures and presents possible solutions to cope the problem of memory organization and data structure. By the example of the MORPHEUS heterogeneous platform, the discussion follows the complete design cycle, starting from decision making and justification, until hardware realization. Particular emphasis is made on the methods to support high system performance, meet application requirements, and provide a user-friendly programmer interface. As a result, the research introduces a complete heterogeneous platform enhanced with a hierarchical memory organization, which copes with its task by means of separating computation from communication, providing reconfigurable engines with computation and configuration data, and unification of heterogeneous computational devices using local storage buffers. It is distinguished from the related solutions by distributed data-flow organization, specifically engineered mechanisms to operate with data on local domains, particular communication infrastructure based on Network-on-Chip, and thorough methods to prevent computation and communication stalls. In addition, a novel advanced technique to accelerate memory access was developed and implemented.

The value of a small microkernel for dreamy memory and the RAMpage memory hierarchy

This paper explores potential for the RAMpage memory hierarchy to use a microkernel with a small memory footprint, in a specialized cache-speed static RAM (tightly-coupled memory, TCM). Dreamy memory is DRAM kept in low-power mode, unless referenced. Simulations show that a small microkernel suits RAMpage well, in that it achieves significantly better speed and energy gains than a standard hierarchy from adding TCM. RAMpage, in its best 128KB L2 case, gained 11% speed using TCM, and reduced energy 14%. Equivalent conventional hierarchy gains were under 1%. While 1MB L2 was significantly faster against lower-energy cases for the smaller L2, the larger SRAM's energy does not justify the speed gain. Using a 128KB L2 cache in a conventional architecture resulted in a best-case overall run time of 2.58s, compared with the best dreamy mode run time (RAMpage without context switches on misses) of 3.34s, a speed penalty of 29%. Energy in the fastest 128KB L2 case was 2.18J vs. 1.50J, a reduction of 31%. The same RAMpage configuration without dreamy mode took 2.83s as simulated, and used 2.39J, an acceptable trade-off (penalty under 10%) for being able to switch easily to a lower-energy mode.

Algorithm-oriented design of efficient many-core architectures applied to dense matrix multiplication

Recent integrated circuit technologies have opened the possibility to design parallel architectures with hundreds of cores on a single chip. The design space of these parallel architectures is huge with many architectural options. Exploring the design space gets even more difficult if, beyond performance and area, we also consider extra metrics like performance and area efficiency, where the designer tries to design the architecture with the best performance per chip area and the best sustainable performance. In this paper we present an algorithm-oriented approach to design a many-core architecture. Instead of doing the design space exploration of the many core architecture based on the experimental execution results of a particular benchmark of algorithms, our approach is to make a formal analysis of the algorithms considering the main architectural aspects and to determine how each particular architectural aspect is related to the performance of the architecture when running an algorithm or set of algorithms. The architectural aspects considered include the number of cores, the local memory available in each core, the communication bandwidth between the many-core architecture and the external memory and the memory hierarchy. To exemplify the approach we did a theoretical analysis of a dense matrix multiplication algorithm and determined an equation that relates the number of execution cycles with the architectural parameters. Based on this equation a many-core architecture has been designed. The results obtained indicate that a 100 mm(2) integrated circuit design of the proposed architecture, using a 65 nm technology, is able to achieve 464 GFLOPs (double precision floating-point) for a memory bandwidth of 16 GB/s. This corresponds to a performance efficiency of 71 %. Considering a 45 nm technology, a 100 mm(2) chip attains 833 GFLOPs which corresponds to 84 % of peak performance These figures are better than those obtained by previous many-core architectures, except for the area efficiency which is limited by the lower memory bandwidth considered. The results achieved are also better than those of previous state-of-the-art many-cores architectures designed specifically to achieve high performance for matrix multiplication.

Reducing main memory access latency through SDRAM address mapping techniques and access reordering mechanisms

As the performance gap between microprocessors and memory continues to increase, main memory accesses result in long latencies which become a factor limiting system performance. Previous studies show that main memory access streams contain significant localities and SDRAM devices provide parallelism through multiple banks and channels. These locality and parallelism have not been exploited thoroughly by conventional memory controllers. In this thesis, SDRAM address mapping techniques and memory access reordering mechanisms are studied and applied to memory controller design with the goal of reducing observed main memory access latency. The proposed bit-reversal address mapping attempts to distribute main memory accesses evenly in the SDRAM address space to enable bank parallelism. As memory accesses to unique banks are interleaved, the access latencies are partially hidden and therefore reduced. With the consideration of cache conflict misses, bit-reversal address mapping is able to direct potential row conflicts to different banks, further improving the performance. The proposed burst scheduling is a novel access reordering mechanism, which creates bursts by clustering accesses directed to the same rows of the same banks. Subjected to a threshold, reads are allowed to preempt writes and qualified writes are piggybacked at the end of the bursts. A sophisticated access scheduler selects accesses based on priorities and interleaves accesses to maximize the SDRAM data bus utilization. Consequentially burst scheduling reduces row conflict rate, increasing and exploiting the available row locality. Using a revised SimpleScalar and M5 simulator, both techniques are evaluated and compared with existing academic and industrial solutions. With SPEC CPU2000 benchmarks, bit-reversal reduces the execution time by 14% on average over traditional page interleaving address mapping. Burst scheduling also achieves a 15% reduction in execution time over conventional bank in order scheduling. Working constructively together, bit-reversal and burst scheduling successfully achieve a 19% speedup across simulated benchmarks.

GreenDisc: a HW/SW energy optimization framework in globally distributed computation

In recent future, wireless sensor networks (WSNs) will experience a broad high-scale deployment (millions of nodes in the national area) with multiple information sources per node, and with very specific requirements for signal processing. In parallel, the broad range deployment of WSNs facilitates the definition and execution of ambitious studies, with a large input data set and high computational complexity. These computation resources, very often heterogeneous and driven on-demand, can only be satisfied by high-performance Data Centers (DCs). The high economical and environmental impact of the energy consumption in DCs requires aggressive energy optimization policies. These policies have been already detected but not successfully proposed. In this context, this paper shows the following on-going research lines and obtained results. In the field of WSNs: energy optimization in the processing nodes from different abstraction levels, including reconfigurable application specific architectures, efficient customization of the memory hierarchy, energy-aware management of the wireless interface, and design automation for signal processing applications. In the field of DCs: energy-optimal workload assignment policies in heterogeneous DCs, resource management policies with energy consciousness, and efficient cooling mechanisms that will cooperate in the minimization of the electricity bill of the DCs that process the data provided by the WSNs.

GreenDisc: a HW/SW energy optimization framework in globally distributed computation

In recent future, wireless sensor networks ({WSNs}) will experience a broad high-scale deployment (millions of nodes in the national area) with multiple information sources per node, and with very specific requirements for signal processing. In parallel, the broad range deployment of {WSNs} facilitates the definition and execution of ambitious studies, with a large input data set and high computational complexity. These computation resources, very often heterogeneous and driven on-demand, can only be satisfied by high-performance Data Centers ({DCs}). The high economical and environmental impact of the energy consumption in {DCs} requires aggressive energy optimization policies. These policies have been already detected but not successfully proposed. In this context, this paper shows the following on-going research lines and obtained results. In the field of {WSNs}: energy optimization in the processing nodes from different abstraction levels, including reconfigurable application specific architectures, efficient customization of the memory hierarchy, energy-aware management of the wireless interface, and design automation for signal processing applications. In the field of {DCs}: energy-optimal workload assignment policies in heterogeneous {DCs}, resource management policies with energy consciousness, and efficient cooling mechanisms that will cooperate in the minimization of the electricity bill of the DCs that process the data provided by the WSNs.

Optimización del framework de paralelización de tareas COMPSs

El propósito de este trabajo es optimizar el sistema de gestión de workflows COMPSs caracterizando el comportamiento de diferentes dispositivos de memoria a nivel de consumo energético y tiempo de ejecución. Para llevar a cabo este propósito, se ha implementado un servicio de caché para COMPSs para conseguir que sea consciente de la jerarquía de memoria y se han realizado múltiples experimentos para caracterizar los dispositivos de memoria y las mejoras en el rendimiento.

Discrete wavelet transform signal analyzer

This paper addresses the problem of processing biological data, such as cardiac beats in the audio and ultrasonic range, and on calculating wavelet coefficients in real time, with the processor clock running at a frequency of present application-specified integrated circuits and field programmable gate array. The parallel filter architecture for discrete wavelet transform (DWT) has been improved, calculating the wavelet coefficients in real time with hardware reduced up to 60%. The new architecture, which also processes inverse DWT, is implemented with the Radix-2 or the Booth-Wallace constant multipliers. One integrated circuit signal analyzer in the ultrasonic range, including series memory register banks, is presented. © 2007 IEEE.

The rapid screen of concussion: An evaluation of the non-word repetition test for use in mTBI research

Primary objectives: (1) To investigate the Nonword Repetition test (NWR) as an index of sub-vocal rehearsal deficits after mild traumatic brain injury (mTBI); (2) to assess the reliability, validity and sensitivity of the NWR; and (3) to compare the NWR to more sensitive tests of verbal memory. Research design: An independent groups design. Methods and procedures: Study 1 administered the NWR to 46 mTBI and 61 uninjured controls with the Rapid Screen of Concussion (RSC). Study 2 compared mTBI, orthopaedic and uninjured participants on the NWR and the Hopkins Verbal Learning Test (HVLT-R). Main outcomes and results: The NWR did not improve the diagnostic accuracy of the RSC. However, it is reliable and indexes sub-vocal rehearsal speed. These findings provide evidence that although the current form of the NWR lacks sensitivity to the impact of mTBI, the development of a more sensitive test of sub-vocal rehearsal deficits following mTBI is warranted.

Improving caches in consolidated environments

Memory (cache, DRAM, and disk) is in charge of providing data and instructions to a computer's processor. In order to maximize performance, the speeds of the memory and the processor should be equal. However, using memory that always match the speed of the processor is prohibitively expensive. Computer hardware designers have managed to drastically lower the cost of the system with the use of memory caches by sacrificing some performance. A cache is a small piece of fast memory that stores popular data so it can be accessed faster. Modern computers have evolved into a hierarchy of caches, where a memory level is the cache for a larger and slower memory level immediately below it. Thus, by using caches, manufacturers are able to store terabytes of data at the cost of cheapest memory while achieving speeds close to the speed of the fastest one.^ The most important decision about managing a cache is what data to store in it. Failing to make good decisions can lead to performance overheads and over-provisioning. Surprisingly, caches choose data to store based on policies that have not changed in principle for decades. However, computing paradigms have changed radically leading to two noticeably different trends. First, caches are now consolidated across hundreds to even thousands of processes. And second, caching is being employed at new levels of the storage hierarchy due to the availability of high-performance flash-based persistent media. This brings four problems. First, as the workloads sharing a cache increase, it is more likely that they contain duplicated data. Second, consolidation creates contention for caches, and if not managed carefully, it translates to wasted space and sub-optimal performance. Third, as contented caches are shared by more workloads, administrators need to carefully estimate specific per-workload requirements across the entire memory hierarchy in order to meet per-workload performance goals. And finally, current cache write policies are unable to simultaneously provide performance and consistency guarantees for the new levels of the storage hierarchy.^ We addressed these problems by modeling their impact and by proposing solutions for each of them. First, we measured and modeled the amount of duplication at the buffer cache level and contention in real production systems. Second, we created a unified model of workload cache usage under contention to be used by administrators for provisioning, or by process schedulers to decide what processes to run together. Third, we proposed methods for removing cache duplication and to eliminate wasted space because of contention for space. And finally, we proposed a technique to improve the consistency guarantees of write-back caches while preserving their performance benefits.^

Supercomputación gráfica aplicada al análisis de imágenes cerebrales con niftyreg

Abstract: Medical image processing in general and brain image processing in particular are computationally intensive tasks. Luckily, their use can be liberalized by means of techniques such as GPU programming. In this article we study NiftyReg, a brain image processing library with a GPU implementation using CUDA, and analyse different possible ways of further optimising the existing codes. We will focus on fully using the memory hierarchy and on exploiting the computational power of the CPU. The ideas that lead us towards the different attempts to change and optimize the code will be shown as hypotheses, which we will then test empirically using the results obtained from running the application. Finally, for each set of related optimizations we will study the validity of the obtained results in terms of both performance and the accuracy of the resulting images.