Citation Context

... work on a new nameless-write SSD interface and virtualized flash storage for file systems have argued for removing the costs of indirection within SSDs by exposing physical flash addresses to the OS =-=[41]-=-, providing caching support [28], and completely delegating block allocation to the SSD [21]. Similar to these systems, FlashTier unifies multiple levels of address space, and provides more control ov...

Citation Context

...llection [7, 16, 17]), and extending their interface (e.g., caching operations [32]). However, most research looking at internal SSD mechanisms relies on simulation rather than direct experimentation =-=[2, 27, 32, 36]-=-. Thus, there is little known about real-world implementation trade-offs relevant to SSD design, such as the cost of changing their command interface. Most such knowledge has remained the intellectual...

Citation Context

...s each layer has limited visibility into the other, and little means to express intent. Current approaches include co-designing a file system with the FTL [15], delegating block allocation to the FTL =-=[9, 29]-=-, and adapting file system layout policies to inferred FTL behavior [8]. This paper argues for an approach that encapsulates these media-specific heuristics from API components, while exposing suffici...

Citation Context

... instance, an SSD can support complex atomic operations [10, 32, 35], native caching operations [5, 38], a large, sparse storage address space [16], delegating storage allocation decisions to the SSD =-=[47]-=-, and offloading file system permission checks to hardware [8]. These new interfaces allow applications to leverage SSDs’ low latency, ample internal bandwidth, and on-board computational resources, a...

Citation Context

... As no extra slot copies are needed, the block erase count is reduced from 2,440 to 1,952, showing its effectiveness. 5 Other Related Work Both flash memory and key-value systems are hot research topics. Due to space constraint, we only present the most related work here. The effect of redundant operations in software layers has been studied before. For example, Yang et al. have pointed out that redundant structures and functions in multiple layers of logs could result in negative effects on flash devices, such as increased write pressure, and should be carefully handled [14]. Nameless Writes [15] presents a new device interface to allow the device to directly control block allocation. FSDV [16] attempts to remove the mapping redundancy by directly storing physical flash addresses in the file systems. Our solution shares the same de-indirection principle as these prior studies. Another set of related work is flash-based key-value stores. For example, SILT [9] optimizes the key-value store with three basic stores for high memory efficiency. NVMKV [10] is a light-weight keyvalue store that builds upon an enhanced FTL. NVMKV extends FTL primitives for simplifying the key-value store desig...

Citation Context

...h each array increases significantly, the processing and reduction of the data before delivering it becomes feasible. Currently missing is an interface which allows applications to easily move data processing to the disks and disk arrays. Knowledge about the application is lost at the storage interface boundary. Read and write optimizations, e.g., for database applications [25], cannot be transferred to the devices. Current approaches try to move the interface into the operating system, to allow it to attach additional semantic information, and to optimize storage accesses for each technology [1, 53], but do not yet deliver the opportunity to directly transform the data. I4: Non-volatile memory technologies Scalability is also limited by the performance of individual drives. While the access speed to inmemory databases only incurs the latency of main memory, object-based disks are still restricted by the underlying magnetic disks and flash. Non-volatile memory technologies (NVRAM) like phase-changed RAM, MRAM, or RRAM will significantly shrink the current performance gap between standard DRAM and storage and reduce access times down to 10 ns. This helps fulfilling R5 and building an econo...

Citation Context

...ddress remapping allows FTLs to improve performance and implement wear leveling, remapping is also often done in the operating system to improve performance for HDDs. A system such as Nameless Writes =-=[13, 83]-=- which integrates the SSD and OS algorithms alleviates these problems. Divination of Applications’ Needs Many of the hybrid designs (described in Chapter 4) improve their efficiency by predicting the ...