Prose I/o Implementation

ions seem to be simultaneously the principal goal and bane of modern computer science[20]. This is particularly true in modern systems in which the system software stack has become a monumental tower of babel. At its foundation we have the virtualization engine or hypervisor, multiplexing system resources between multiple logical partitions. Each of these logical partitions is running an operating system which facilitates sharing of resources between distinct processes and also provides system service abstractions for file systems, networking, and process control. On top of the operating system we have a set of facilities for distributed computing such as a message-passing-toolkit (MPI)[21] or the OpenMP[22] libraries. On top of the distributed system layer we have virtual machines[23][24][25], providing yet another layer of abstraction, system services, and resource scheduling. On server systems a middleware layer [26] runs on top of the virtual machine providing another duplicate set of system services and yet another layer of abstraction. On top of the middleware runs the application, which itself may be an interpreter or execution engine (such as one for an XML business process) [27]. Figure 1. The modern software stack This is a worst case view of the modern software stack, but it is a surprisingly prevalent one within service-oriented-architectures [28]. Each layer has its own resource allocation and arbitration logic. For example, there are separate timesharing schedulers present in each layer. Each layer of abstraction hides its implementation details from the layer above it, making coordinated resource management impossible. What results is a debugging and performance tuning nightmare, with each layer adding both overhead and interference to the other layers. Additionally, the various services provided by each layer are usually immutable. Developers or users can not easily override the scheduler or memory allocation mechanisms of underlying layers. This is particularly true within the operating system layer, where privileged accounts are necessary to even interact with system services. Microkernels [29][30] attempted to solve this problem by moving operating system services into user-space. However, provisioning and configuration of these services was still a privileged operation. Plan 9's [31] user-composable name spaces provides a different approach allowing user replacement of typical system services -but doesn't allow user control of more complex system services such as memory allocation or scheduling. IBM's research operating system K42[32], allowed hot-swapping of these lower level kernel services, but provided a Linux API/ABI abstraction which ultimately limited its flexibility. The problem was probably best approached by MIT's Exokernel[5] research operating system which limited itself to straightforward multiplexing of hardware resources, and relied upon optional libraries (libOS) to provide familiar operating system abstractions. Developer's could choose which library instance to link their application against, or "roll their own" service or abstraction. This allows for an unprecedented degree of flexibility in the implementation of operating system abstractions, allowing application specific customizations and optimizations. Microkernels, Plan 9, K42, and Exokernels never saw mainstream use principally because of their limited hardware platform support and lack of ability to run popular end-user applications. What seemed most desirable was a mainstream environment which could be used for the user environment and which could be "pushed out of the way" for execution of specialized applications which have their own system abstractions and services. Having the ability to leverage certain services and resources from the mainstream operating system would give developers the ability to focus effort on the handling of resources and services most critical to their application. Such a customizable environment would be particularly attractive within specialized workload domains such as high-performance computing and real-time systems.