Citation Context

...lays in data centers that host delay-sensitive services. This work includes new network architectures [14, 2, 18], new transport protocols [3, 31], and techniques to rearrange computation and storage =-=[1, 23]-=-. None of these proposals is concerned with managing end-to-end deadlines. Time synchronization: A number of innovative proposals for time synchronization in various domains, such as the Internet [26,...

Citation Context

...ork has shown that 8MB random disk reads can achieve approximately 88% of the throughput of sequential reads, and that smaller random reads cause significant performance degradation due to disk seeks =-=[23]-=-. Thus, as long as input data is stored on disk, we require that tasks read at least 8MB of input data. Data-parallel workloads are commonly I/O bound, so we expect task runtime to be dominated by tim...

Citation Context

...e the speed and sophistication of large-scale dataparallel processing systems. Practitioners and researchers have built a wide array of programming frameworks [34– 36, 42, 53, 54] and storage systems =-=[14, 15, 25, 26, 39]-=- tailored to a variety of workloads. As the performance of many of these systems is I/O bound, traditional means of improving their speed is to cache data into memory [12]. While caching can dramatica...

Citation Context

...ey do not support storage heterogeneity. In the distributed setting, Zebra [15], GPFS [31], and Panasas [36], offer techniques to improve read and write throughput. Similarly, Flat Datacenter Storage =-=[25]-=- that employs an advanced network topology, and Camdoop [8] that uses a directly-connected network topology, argue the need for better provisioning and utilization of the disk bandwidth in the modern ...

Citation Context

...cability of our approach. Exalt is primarily designed to evaluate I/O-intensive applications like distributed file systems storing large files [7, 22] or key-value stores with relatively large values =-=[3, 17]-=-. Applications that are not I/O-intensive or store small values can not benefit significantly from Tardis, as they gain little by compressing data. In Section 6.2 we explore in more detail how the siz...

Citation Context

...te I/Os that are transient to support the execution (e.g. shuffling, spilling and merging). The ratio between the size of map input read and reduce output write is often described as reduction factor =-=[23]-=-. It indicates the level of condensation as data is processed throughout the framework, relative to its original input. When reduction factor is high, the amount of intermediate I/Os are generally sma...

Citation Context

...orts on both computation and storage layers to make big data processing faster. Practitioners and researchers have built a wide array of programming frameworks [10, 18, 11, 8, 17] and storage systems =-=[6, 13, 2]-=- tailored to different workloads. A key observation that has been made to speed up processing is that working sets in most datacenters are comparatively small, which allows hot data to be kept in memo...

Citation Context

...or a unique purpose, asynchronously uploading or downloading objects to backend services. While algorithms for distributing or replicating objects have also been proposed and explored by past systems =-=[6, 24, 25]-=-, MetaSync devised our replication scheme for deterministic way to minimize sharing, and also to satisfy space restrictions of multiple backends. A major line of work, starting with SUNDR [18] but car...

Citation Context

...ays in data centers that host delay-sensitive services. This work includes new network architectures [40, 9, 51], new transport protocols [10, 78], and techniques to rearrange computation and storage =-=[4, 65]-=-. None of these proposals is concerned with managing end-to-end deadlines. Time synchronization: A number of innovative proposals for time synchronization in various domains, such as the Internet [70,...