Next Generation Software: SmartApps: Smart Applications for Heterogeneous Computing

点击翻译按钮获取中文摘要

EIA-9975018Rauchwerger, LawrenceCISE Next Generation Software(NG)Texas Engineering Experimental StationSmartApps: Smart Applications for Heterogeneous ComputingThis proposal proposes to generate SMART APPLICATION code that uses information about its input data, operating system services and architecture capability, to adaptively reconfigure the application and the OS/hardware platform. To enable very aggressive optimizations, we propose to build a two-stage compiler which has a static and run-time component. The run-time component is extremely fast, works in parallel on the target machine and can read all relevant input data, and the operating system and architectural capabilities. Furthermore, to adapt the performance to variable program inputs during execution, we propose to use novel speculative run-time adaptive optimization techniques (for example, the PI's speculative parallelization and a feedback loop of performance evaluator, predictor and optimizer inputs. All this support will be enabled by a run-time library and the necessary compiler-embedded code in the executable. With this approach, the application can adapt itself and its environment to its needs by issuing appropriate commands. For instance, it is able to control data placement by the OS, control the I/O subsystem, modify coherence protocols and establish network connections. Moreover, smart applications are able to respond to the failures and large performance variations that distributed systems suffer.