Citation Context

...ension to memory for random read caching. Windows gradually fills the flash cache with data and then services random reads from that cache, when doing so improves performance. Recently, Koller et al. =-=[11]-=- experimented with a range of more sophisticated writeback policies for a flash cache. They found (as we did) that synchronous writethrough all the way to disk is slow. Their work is otherwise complem...

Citation Context

... way to reduce BCH’s false positives, as opposed to our replacement of BCH for clean pages. Yang et al. [45] propose to improve SSD cache endurance via reducing media writes and erases. Koller et al. =-=[22]-=- present a study discussing write policies and consistency problems of SSD cache deployed in networked environment. More recently, Holland et al. [18] explore the design space of flash as a cache in t...

Citation Context

...., minimizing average VM latency or meeting a specific VM’s latency target). Host-side SSDs are a recent augmentation to disk-based storage stacks that are becoming increasingly mainstream [6], [16], =-=[20]-=-, [30], [42]. When used as a cache, the SSD space can be dynamically partitioned across workloads to meet per-workload performance goals. Current cache partitioning techniques optimize aggregate metri...

Citation Context

...e of a nondatapath cache. Host-side flash caches are attractive because they can reduce the demands placed on network storage, speed up I/O performance, and provide I/O latency and throughput control =-=[1, 6, 7]-=-. A cache update is composed of two operations: an eviction and an insertion. Performance wise, evictions can be detrimental. In case of a host-side flash cache, a dirty item chosen for eviction needs...

Citation Context

... fetch rates also continue to increase. A host of flash-based hybrid memory systems and storage caching and tiering systems have been inspired, and find relevance in practice, because of these trends =-=[3, 4, 13, 16, 17, 18, 22, 24, 35, 40, 45, 55, 57]-=-. On the personal computing end, newer data intensive desktop/laptop applications place greater I/O demands [20]. In mobile systems, page fetches have been found to affect the performance of the data-...

Citation Context

...r [7]. On the other hand, a file-based interface provides the “killer” advantage of configurable guest isolation and sharing support at fine granularity [7]–[9], [14], [15]. The file-based interface also facilitates semantics awareness, which strengthens the consistency and fault isolation of filesystems, and enables several performance optimizations. For instance, the treatment of metadata as data by the block-based interface makes a guest susceptible to corruption from system crashes, unless all writes are treated as synchronous below the guest (e.g., with write-through caching) [12], [13], [16]. Also, filesystem nesting through a block-based interface may lead to performance degradation, e.g., in write-dominated or latencysensitive workloads [12], [13]. Wide compatibility across different backend storage systems and frontend guest operating systems makes the block-based978-1-4799-5671-5/14/$31.00 © 2014 IEEE interface a preferable choice [13]. Instead, several manageability benefits are currently provided by both interfaces. Versioning is supported at coarse granularity in virtual disks, or fine granularity with individual files [7]. Both virtual disks and distributed filesystems ca...

Citation Context

... Koller et al. also advocated the importance of write-back caching and studied new ordered and journaled write-back policies for flash caches, in order to improve the consistency of cached dirty data =-=[23]-=-. This paper complements the previous work by further studying the performance impact of write-back caching to both storage client and server using real workloads and proposing a new cache-optimized R...

Citation Context

...torage-level inconsistencies after a cache failure [6]. A. Consistent, Block-Level Write-Back Caching To solve these issues caused by unordered write-back policy, two solutions were proposed recently =-=[12]-=-. The first technique, referred to as Ordered Write Back, is based on the simple idea that consistency at the networked storage system can be maintained by evicting blocks in the same order in which t...