روش مدل شبیه سازی جعبه سیاه جدید برای پیش بینی عملکرد و مصرف انرژی در دستگاه های ذخیره سازی کالا

Traditional approaches for storage devices simulation have been based on detailed and analytic models. However, analytic models are difficult to obtain and detailed models require a high computational cost which may be not affordable for large scale simulations (e.g. detailed data center simulations). In current systems like large clusters, grids, or clouds, performance and energy studies are critical, and fast simulations take an important role on them. A different approach is the black-box statistical modeling, where the storage device, its interface, and the interconnection mechanisms are modeled as a single stochastic process, defining the request response time as a random variable with an unknown distribution. A random variate generator can be built and integrated into a bigger simulation model. This approach allows to generate a simulation model for both real and synthetic complex workloads. This article describes a novel methodology that aims to build fast simulation models for storage devices. Our method uses as starting point a workload and produces a random variate generator which can be easily integrated into large scale simulation models. A comparison between our variate generator and the widely known simulation tool DiskSim, shows that our variate generator is faster, and can be as accurate as DiskSim for both performance and energy consumption predictions.

As demonstrated by the successful emergence of the Green500 [8] list, which provides a ranking of the most energy-efficient supercomputers in the world, energy has become as significant as performance. Consequently, the performance-per-watt has been established as a new metric to evaluate systems. Many researchers have shown interest in identifying in which cases there is room for improvement in the context of power efficiency. As a result, a lot of inefficiencies in relation to energy have been identified. Research works show that a CPU resting in an idle state reaches about 50% of peak power consumption [10]. Storage subsystems alone represent roughly 10–25% of the power consumed by the data center [14]. High power consumption in storage systems is expected, as an idle machine with one processor and two disks can easily spend as much power on disks as on processors [22]. Storage consumption can become a greater problem in storage subsystems where the average number of disks per machine is in the dozens. In this way, simulation techniques are commonly used for both performance and energy consumption evaluation of many applications and systems. In order to obtain realistic results, data access to either a file system or database management system cannot be ignored. However, a key element in the data access studies is the storage device simulation model. A storage device simulation model accepts as input the parameters of an application workload (as a flow of device requests) and generates a performance metric prediction. The output performance measurement may be a general performance metric such as average bandwidth, throughput, and latency. Such metrics give an idea of the global performance of the device. However, if the device simulation model is integrated on top of a large system model, such as file systems or database managers, a detailed metric as request’s response time is needed. Even more, disk’s manufactures provide response time as an average value however, these metrics are not suitable for detailed simulations.

Providing efficient and power-aware storage solutions is currently an active research field. There is a growing number of optimizations and solutions which deal with the problem of achieving energy efficiency and performance, however there is a lack of tools which allow evaluate then. In this paper we have described a methodology used to build fast and accurate black-box simulation models for hard disk drives. The presented solution is based on probability distributions and can be used for both synthetic and real traces. The methodology includes a generic measuring service time tool. Our approach is based on a multiple real trace repository and may be accurate when the traces to predict have similar characteristics as one of the traces, for which the model has been constructed. The model can also be based on a single synthetic trace. This trace is supposed to cover a wide range of characteristics from other traces, including several access patterns, common in real disks. Additionally, this approach allows to calculate queuing times on the fly on prediction stage, making it more versatile. Our solution has been compared with DiskSim in terms of accuracy and speed-up, showing that BBMP is two orders of magnitude faster than DiskSim for most of the experiments. It has also been compared with DiskSim in terms of energy consumption estimations, and we concluded that it keeps being accurate in that respect.