Recently, ultra-low latency flash storage devices such as Z-SSD and Optane SSD were introduced with the significant technological improvement in the storage devices which provide much faster response time than today's other NVMe SSDs. With such ultra-low latency, $10{\mu}s$, storage devices the cost of context switch could be an overhead during interrupt-driven I/O completion process. As an interrupt-driven I/O completion process could bring an interrupt handling overhead, polling or hybrid-polling for the I/O completion is known to perform better. In this paper, we analyze tail latency problem in a polling process caused by process scheduling in data center environment where multiple applications run simultaneously under one system and we introduce our adaptive polling selection technique which dynamically selects efficient processing method between two techniques according to the system's conditions.

As the capacity of SSDs rapidly increases, the amount of DRAM to keep a mapping table size in SSDs becomes very huge. To address a Demand-based FTL (DFTL) scheme that caches part of mapping entries in DRAM is considered to be a feasible alternative. However, owing to its unpredictable behaviors, DFTL fails to provide consistent I/O response times. In this paper, we a) analyze a root cause that results in fluctuation on read latency and b) propose a new demand-based FTL scheme that ensures guaranteed read response time with low write amplification. By preventing mapping evictions while serving reads, the proposed technique guarantees every host read requests to be done in 2 NAND read operations. Moreover, only with 25% of a cache ratio, the proposed scheme improves random write performance and random mixed performance by 1.65x and 1.15x, respectively, over the traditional DFTL.

For decades, in-memory data structures have been designed for DRAM-based main memory that provides symmetric read/write performances and has no limited write endurance. However, such data structures provide sub-optimal performance for NVM as it has different characteristics to DRAM. With this motivation, we rethink a conventional red-black tree in terms of its efficacy under NVM settings. The original red-black tree constantly rebalances sub-trees so as to export fast access time over dataset, but it inevitably increases the write traffic, adversely affecting the performance for NVM with a long write latency and limited endurance. To resolve this problem, we present a variant of the red-black tree called a hierarchical balanced binary search tree. The proposed structure maintains multiple keys in a single node so as to amortize the rebalancing cost. The performance study reveals that the proposed hierarchical binary search tree effectively reduces the write traffic by effectively reaping the high capacity of NVM.

Low power issue is one of the most critical problems in the Internet of Things (IoT), which are powered by battery. To solve this problem, various approaches have been presented so far. In this paper, we propose a method to reduce the power consumption by reducing the numbers of accesses into the flash memory consuming a large amount of power for on-chip software execution. Our approach is based on using cooperative logging structure to distribute the sampling overhead in single sensor node to adjacent nodes in case of rare-event applications. The proposed algorithm to identify event occurrence is newly introduced with negative feedback method by observing difference between past data and recent data coming from the sensor. When an event with need of flash access is determined, the proposed approach only allows access to write the sampled data in flash memory. The proposed event detection algorithm (EDA) result in 30% reduction of power consumption compared to the conventional flash write scheme for all cases of event. The sampled data from the sensor is first traced into the random access memory (RAM), and write access to the flash memory is delayed until the page buffer of the on-chip flash memory controller in the micro controller unit (MCU) is full of the numbers of the traced data, thereby reducing the frequency of accessing flash memory. This technique additionally reduces power consumption by 40% compared to flash-write all data. By sharing the sampling information via LoRa channel, the overhead in sampling data is distributed, to reduce the sampling load on each node, so that the 66% reduction of total power consumption is achieved in several IoT edge nodes by removing the sampling operation of duplicated data.

Multi-queue I/O block layer has been recently employed in Linux kernel to support fast storage devices such as NVMe SSDs, but it lacks differentiated I/O services yet. In this paper, we propose an I/O scheduling scheme that can improve the user responsiveness of foreground processes, which are closely related to user satisfaction. To this end, we redesign the existing multi-queue block layer to classify the I/O requests from foreground processes and schedule them by exploiting the feature of NVMe interface. Experimental results show that latency and launch time of the foreground processes have been significantly improved compared to original Linux kernel.

The IoT-driven large-scaled systems consist of connected things with on-chip executable embedded software. These light-weighted embedded things have limited hardware space, especially small size of on-chip flash memory. In addition, on-chip embedded software in flash memory is not easy to update in runtime to equip with latest services in IoT-driven applications. It is becoming important to develop light-weighted IoT devices with various software in the limited on-chip flash memory. The remote instruction execution in cloud via IoT connectivity enables to provide high performance software execution with unlimited software instruction in cloud and low-power streaming of instruction execution in IoT edge devices. In this paper, we propose a Cloud-IoT asymmetric structure for providing high performance instruction execution in cloud, still low power code executable thing in light-weighted IoT edge environment using remote instruction execution. We propose a simulated approach to determine efficient partitioning of software runtime in cloud and IoT edge. We evaluated the instruction cloudification using remote instruction by determining the execution time by the proposed structure. The cloud-connected instruction set simulator is newly introduced to emulate the behavior of the processor. Experimental results of the cloud-IoT connected software execution using remote instruction showed the feasibility of cloudification of on-chip code flash memory. The simulation environment for cloud-connected code execution successfully emulates architectural operations of on-chip flash memory in cloud so that the various software services in IoT can be accelerated and performed in low-power by cloudification of remote instruction execution. The execution time of the program is reduced by 50% and the memory space is reduced by 24% when the cloud-connected code execution is used.