A well-known problem for large scale cloud applications is how to scale their I/O performance. While next generation storage class memories like phase change memory and Memristors offer potential for high I/O bandwidths, if left unchecked, the raw volumes and rates of I/O already present in cloud can quickly overwhelm future I/O infrastructures. This fact is motivating research on 'data staging' in which I/O and data movement are partly replaced with and/or supplemented by computations that process data before moving it across I/O channels * in situ * to filter or reduce it, to better organize it for subsequent access (e.g., by other applications as in coupled codes), or to analyze it to quickly to derive important insights about the application producing those large data volumes. This paper proposes a technique that uses and exploits 'Active NVRAM' (non volatile memory) for staging I/O. Active NVRAMs are node-local NVRAMs that are embedded with a low power system-on-chip compute element. These active compute elements can be used to operate on output data asynchronously with the tasks performed by computational node elements, to reduce data or to perform some of the data processing required for data analytics before data is moved to longer term storage. The paper further describes the Active NVRAM design, sample ways in which it is used for I/O acceleration, and initial performance results evaluating the opportunities for and limitations of the Active NVRAM approach.
PDF
download
878987797
028-85220240
