University of Delaware Department of Electrical and Computer Engineering Computer Architecture and Parallel Systems Laboratory Collaborative Research: Programming Models and Storage System for High Performance Computation with Many-Core Processors

Future generation HEC architectures and systems built using many-core chips will pose unprecedented challenges for users. A major source of these challenges is the storage system since the available memory and bandwidth per processor core is starting to decline at an alarming rate, with the rapid increase in the number of cores per chip. Data-intensive applications that require large data sets and/or high input/output (I/O) bandwidth will be especially vulnerable to these trends. Unlike previous generations of hardware evolution, this shift in the hardware road-map will have a profound impact on HEC software. Historically, the storage architecture of an HEC system has been constrained to a large degree by the file system interfaces in the underlying Operating System (OS). The basic design of file systems has been largely unchanged for multiple decades, and will no longer suffice as the foundational storage model for many-core HEC systems. Instead, it becomes necessary to explore new approaches to storage systems that reduce OS overhead, exploit multiple levels of storage hierarchy and focus on new techniques for efficient utilization of bandwidth and reduction of average access times across these levels. The specific focus of this proposal is on exploring a new storage model based on write-once tree structures, which is radically different from traditional flat files. The basic unit of data transfer is a chunk, which is intended to be much smaller than data transfer sizes used in traditional storage systems. The write-once property simplifies the memory model for the storage system and obviates the need for complicated consistency protocols. We will explore three programming models for users of the storage system, all of which can inter-operate through shared persistent data: 1) a declarative programming model in which any data structure can be directly made persistent in our storage system, with no programmer intervention, 2) a strongly-typed imperative programming model in which a type system extension will be used to enforce a separation between data structures that can be directly made persistent and those that cannot, and 3) a weakly-typed runtime interface that enables C programs to access our storage system. A compiler with a high-level data flow intermediate representation and lower-level parallel intermediate representation will provide the necessary code generation support, and a runtime system will implements interfaces to the storage system used by compiler-generated code and by the weakly-typed runtime interface. Our proposed research will be evaluated using an experimental testbed that …