Search

Simulation is a fundamental research tool in the computer architecture field. These kinds of tools enable the exploration and evaluation of architectural proposals capturing the most relevant aspects of the highly complex systems under study. Many state-of-the-art simulation tools focus on single-system scenarios, but the scalability required by trending applications has shifted towards distributed computing systems integrated via complex software stacks. Web services with client-server architectures or distributed storage and processing of scale-out data analytics (Big Data) are among the main exponents. The complete simulation of a distributed computer system is the appropriate methodology to conduct accurate evaluations. Unfortunately, this methodology could have a significant impact on the already large computational effort derived from detailed simulation. In this work, we conduct a set of experiments to evaluate this accuracy/cost tradeoff. We measure the error made if client-server applications are evaluated in a single-node environment, as well as the overhead induced by the methodology and simulation tool employed for multi-node simulations. We quantify this error for different micro-architecture components, such as last-level cache and instruction/data TLB. Our findings show that accuracy loss can lead to completely wrong conclusions about the effects of proposed hardware optimizations. Fortunately, our results also demonstrate that the computational overhead of a multi-node simulation framework is affordable, suggesting multi-node simulation as the most appropriate methodology., Comment: 10 pages, 11 figures

This work explores the feasibility of specialized hardware implementing the Cortical Learning Algorithm (CLA) in order to fully exploit its inherent advantages. This algorithm, which is inspired in the current understanding of the mammalian neo-cortex, is the basis of the Hierarchical Temporal Memory (HTM). In contrast to other machine learning (ML) approaches, the structure is not application dependent and relies on fully unsupervised continuous learning. We hypothesize that a hardware implementation will be able not only to extend the already practical uses of these ideas to broader scenarios but also to exploit the hardware-friendly CLA characteristics. The architecture proposed will enable an unfeasible scalability for software solutions and will fully capitalize on one of the many CLA advantages: low computational requirements and reduced storage utilization. Compared to a state-of-the-art CLA software implementation it could be possible to improve by 4 orders of magnitude in performance and up to 8 orders of magnitude in energy efficiency. We propose to use a packet-switched network to tackle this. The paper addresses the fundamental issues of such an approach, proposing solutions to achieve scalable solutions. We will analyze cost and performance when using well-known architecture techniques and tools. The results obtained suggest that even with CMOS technology, under constrained cost, it might be possible to implement a large-scale system. We found that the proposed solutions enable a saving of 90% of the original communication costs running either synthetic or realistic workloads., Comment: Submitted for publication

Chip multiprocessors (CMPs) are ubiquitous in most of today's computing fields. Although they provide noticeable benefits in terms of performance, cost and power efficiency, they also introduce some new issues. In this paper we analyze how the interference from Virtual Private Servers running in other cores is a significant component of performance unpredictability and can threaten the attainment of cloud computing. Even if virtualization is used, the sharing of the on-chip section of the memory hierarchy by different cores makes performance isolation strongly dependent on what is running elsewhere in the system. We will show in three actual computing systems, based on Sun UltraSparc T1, Sun UltraSparc T2 and Intel Xeon processors, how state-of-the-art virtualization techniques are unable to guarantee performance isolation in a representative workload such as SPECweb2005. In an especially conceived near worst-case scenario, it is possible to reduce the performance achieved by a Solaris Zones consolidated server for this suite of benchmarks in a Sun Fire T1000 and a Sun Enterprise T5120 by up to 80%. The performance drop observed by a Xen consolidated server running in a HP Proliant DL160 G5 is almost 45%. For all systems under study, off-chip bandwidth is shown to be the most critical resource.

This document shows the detailed specification of LOCKE coherence protocol for each cache controller, using a table-based technique. This representation provides clear, concise visual information yet includes sufficient detail (e.g., transient states) arguably lacking in the traditional, graphical form of state diagrams., Comment: 3 pages