Mohammed G. Khatib

Optimizing MEMS-based storage devices for mobile battery-powered systems

An emerging storage technology, called MEMS-based storage, promises nonvolatile storage devices with ultrahigh density, high rigidity, a small form factor, and low cost. For these reasons, MEMS-based storage devices are suitable for battery-powered mobile systems such as PDAs. For deployment in such systems, MEMS-based storage devices must consume little energy. This work mainly targets reducing the energy consumption of this class of devices. We derive the operation modes of a MEMS-based storage device and systemically devise a policy in each mode for energy saving. Three types of policies are presented: power management, shutdown, and data-layout policy. Combined, these policies reduce the total energy consumed by a MEMS-based storage device. A MEMS-based storage device that enforces these policies comes close to Flash with respect to energy consumption and response time. However, enhancement on the device level is still needed; we present some suggestions to resolve this issue.

Interposing Flash between Disk and DRAM to Save Energy for Streaming Workloads

In computer systems, the storage hierarchy, composed of a disk drive and a DRAM, is responsible for a large portion of the total energy consumed. This work studies the energy merit of interposing flash memory as a streaming buffer between the disk drive and the DRAM. Doing so, we extend the spin-off period of the disk drive and cut down on the DRAM capacity at the cost of (extra) flash. We study two different streaming applications: mobile multimedia players and media servers. Our simulated results show that for light workloads, a system with a flash as a buffer between the disk and the DRAM consumes up to 40% less energy than the same system without a flash buffer. For heavy workloads savings of at least 30% are possible. We also address the wear- out of flash and present a simple solution to extend its lifetime.

Power management of MEMS-based storage devices for mobile systems

Because of its small form factor, high capacity, and expected low cost, MEMS-based storage is a suitable storage technology for mobile systems. MEMS-based storage devices should also be energy efficient for deployment in mobile systems. The problem is that MEMS-based storage devices are mechanical, and thus consume a large amount of energy when idle. Therefore, a power management (PM) policy is needed that maximizes energy saving while minimizing performance degradation. In this work, we quantitatively demonstrate the optimality of the fixed-timeout PM policy for MEMS-based storage devices. Because the media sled is suspended by springs across the head array in MEMS-based storage devices, we show that these devices (1) lack mechanical startup overhead and (2) exhibit small shutdown overhead. As a result, we show that the combination of a PM policy, that fixes the timeout in the range of 1--10 ms, and a shutdown policy, that exploits the springs, results in a near-optimal energy saving yet at a negligible loss in performance.

Performance Evaluation of Three Architectural Variants for Multi-sled MEMS Storage

MEMS storage technology promises appealing features, such as ultrahigh density and low cost. To stay competitive, however, its performance must improve to fulfil the increasing I/O requirements in first-tier applications. Further, it must be energy-efficient. The maturity of MEMS techniques invites for a change in the basic single-medium architecture of MEMS storage. We propose three MEMS architectures of multiple media sleds. The three architectures vary in their performance and cost, and consume 35\% less energy than the basic single-sled architecture of MEMS storage. We compare their timing performance in mobile, streaming, and database applications, and study their performance scalability.

Policies for probe-wear leveling in MEMS-based storage devices

Probes (or read/write heads) in MEMS-based storage devices are susceptible to wear. We study probe wear, and analyze the causes of probe uneven wear. We show that under real-world traces some probes can wear one order of magnitude faster than other probes leading to premature expiry of some probes. Premature expiry has severe consequences for the reliability, timing performance, energy-efficiency, and the lifetime of MEMS-based storage devices. Therefore, wear-leveling is a must to preclude premature expiry. We discuss how probe wear in MEMS-based storage is different from medium wear in Flash, calling for a different treatment. We present three policies to level probe wear. By simulation against three real-world traces, our work shows that an inevitable trade-off exists between lifetime, timing performance, and energy efficiency. The policies differ in the size of the trade-off. One of the policies maximizes the lifetime, so that it is optimal; and the other two are less optimal, and are used based on the configuration of the device.

Fast configuration of MEMS-based storage devices for streaming applications

An exciting class of storage devices is emerging: the class of Micro-Electro-Mechanical storage Systems (MEMS). Properties of MEMS-based storage devices include high density, small form factor, and low power. The use of this type of devices in mobile infotainment systems, such as video cameras is not at all obvious. We must explore their configuration and assess their benefit with respect to existing devices, such as Flash.