Real-Time Virtual Resources

点击翻译按钮获取中文摘要

Aloysius K. MokCCR-0207853 "REAL-TIME VIRTUAL RESOURCES"As embedded systems become more complex, a typical embedded systemwill probably involve a mix of soft and hard real-time applicationsthat share the same embedded run-time platform. In general, anapplication may consist of one or more tasks. For example, on ahand-held sensor-computer, a user may want to fix his own locationwith respect to a target landmark and also be warned of the powerconsumption status of the sensor-computer. While the power consumptionwarning task may have very loose timing constraints, the locationfixing task has fairly stringent timing and other QOS (quality ofservice) requirements. Ideally, an application developer should beable to write his application program as if it were running on adedicated computer and not have to worry about interference fromother applications. This illusion can be maintained if globalknowledge of the QOS requirements of all the applications isaccessible to the operating system scheduler, so that a globalschedulability analysis can be performed to ensure that every taskmeets its QOS requirements. However, this is not possible in theenvironment of "open systems" where the operating system may notknow about the timing requirements of all the application tasks,as is the case with most commercial embedded and real-time kernels.A simplistic solution is to assign a higher priority to timing-critical applications and let the non-real-time applications runonly when the timing-critical ones are finished. This approach isviable only for the simplest embedded systems where there is onlyone real-time task and there is no interaction among the real-timeand the non-real-time tasks. If all the applications are to variousdegrees timing-critical (hard or soft), then the distinction betweenreal-time and non-real-time applications becomes blurry. In thesensor-computer example, if the power consumption warning task isassigned a low priority because of its laxer timing constraint, thenit is possible that power may run out while the high-priority locationfixing task is monopolizing the CPU. This may in turn cause a sensorreset that may affect the validity of the location fixing result.The goal of this project is to enable the design of robust embeddedreal-time systems that must function in an open systems environment.The innovation we are investigating is an elegant software abstractioncalled RTVR (real-time virtual resource). This abstraction allowsapplication programmers to design embedded real-time systems as ifeach application had exclusive access to a set of dedicated physicalresources which provide service at a roughly constant rate. The keyto the RTVR concept is that of a "delay bound" parameter on the rateof service provision. The delay bound specification enables aprogrammer to specify the jitter allowance of his/her application.As a result, we can simplify the verification of the timing correctnessof individual applications as if there is no timing interferenceamong application tasks. The RTVT concept also simplifies the issueof guaranteeing the timing/QOS requirements of a mixture of hardand soft real-time applications. To realize this abstraction, wemust make major advances in both theory and engineering in operatingsystem design and real-time scheduling theory. Of particular interestto this project are the issues of partition composition and thehierarchical decomposition of multiple real-time virtual resources.