Vote for peace: implementation and performance of a parallel operating system

ion from the physical network interface is handled by the cluster bus (CLUB). (This terminology comes from the Suprenum architecture, where a cluster bus interconnected up to 20 nodes to build a cluster. A second-level network system interconnected up to 16 clusters. The low-level communication protocol was known as the cluster bus driver.) Thus, the bottom layer of the communication system encapsulates the network device and physically attaches the nucleus to the network. This layer implements the network device driver. CLUB provides the view of an abstract network device that can have several physical representations. The CLUB abstraction makes COSY independent from the network device actually used, whether this device is a physical or a logical one. Thus, CLUB supports the portability of COSY protocols. DUAL-PROCESSOR-NODE CONFIGURATIONS The basic idea behind the Manna dual-processor-node architecture is to have one processor in charge of application program processing and to use the second processor for global communication (internode message passing). The Intel Paragon, for example, implements a similar node architecture. This technology’s main aim is to provide architectural support to minimize the message startup time and the message latency. However, all these measures are of little value if the operating-system architecture for such a hardware organization is inappropriate. For example, both OSF/1 AD10 and Puma11 are parallel operating systems for the Paragon machine. One of the main reasons that Puma outperforms OSF/1 AD is that the former has been specifically designed to operate in a distributed-memory parallel-computer environment, whereas the latter is mainly a port of a microkernel-based distributed (timesharing) operating system. Peace falls into the same category as Puma. The two processors of a Manna node are fully software-programmable and can be configured in various ways. Singleand multiprocessor configurations are equally supported. Depending on the application demands, symmetric or asymmetric multiprocessing might come into play. For example, one processor might play the role of an application processor, and the other processor might play the role of a communication processor. The idea of the CP is to relieve the AP of all functions necessary for driving the networkwide communication protocol. In the following paragraphs, we discuss in more detail the two main Manna node configurations that the Peace kernel supports. Peace supports other configurations, such as symmetric multiprocessing; however, they have not reached the same importance for the Manna applications, so we don’t describe them here.