Chapter 26 Extensible systems

Exercises

26-1 To what extent do you agree that the problems involved in designing a successful microkernel-based single-machine operating system mirror those involved in designing distributed systems in general? In what ways do the two problems differ?

Both problems require a consideration of, amongst other things, communication, naming, security and the management of component failures. In the case of a single-machine operating system, the range of failures may be simplified if the local communication mechanisms can be assumed to be reliable (and are likely to be substantially more lightweight than an RPC). There is no analogy for a ‘network partition’.

26-2 Describe a number of optimisations which are enabled by demultiplexing network packets into paths. What are the disadvantages?

Essentially, if the series of modules which will process data is known ahead of time then inter-module optimisation may be possible – for instance by extracting the fast-path code (i.e. the code normally executed when handling the majority of valid data packets) from each module and collecting that together in a single function.

26-3 The Nemesis operating system (Section 26.5) uses a single-virtual address space process model in which all processes execute within the same virtual address space, albeit with different virtual memory protection settings. Assess the advantages and disadvantages of this approach. What impact does it have on providing binary-compatibility for UNIX applications?

A major advantage of this approach is that pointers can be directly shared between processes whenever permitted by the virtual memory protection settings – the ideas of translation and protection have been separated. This aids the use of shared memory communication structures such as the Rbufs channel. One particular difficulty posed for supporting UNIX applications is how to support fork-based process creation. Normally each UNIX process follows the same conventions as to how to lay out its virtual address space.

26-4 The scheduling scheme used in Nemesis was presented in Section 26.5 for uniprocessor systems. How would you extend it to allow a multi-threaded process to make use of a machine with several CPUs?

A basic scheme could be developed by performing separate activations of the same process on each of the CPUs it is executed on. Either the OS, or the user-level thread scheduler, would have to enforce appropriate concurrency control in order to avoid placing the same thread on several CPUs at the same time. One approach would be for the OS to prevent concurrent execution of a process’s activation handler on more than one CPU. This would also simplify the management of ‘resume’ activation slots.

26-5 A JVM is to execute a number of threads performing separate tasks and to provide resource-isolation between them. Describe, for three different resources, why such execution requires support from the JVM implementer.

An application running over the JVM could loop endlessly at a high priority, or it could loop endlessly creating threads or it could loop endlessly allocating objects and causing work for the garbage collector.