Discussion Questions for 12/06

Distributed Synchronization: Transactions

• Would you explain transactions in class for us? (Nick)
• On page 725, Nutt says that transactions "are useful in systems that are susceptible to crash or in which the consequences of an unfortuitous crash are catastrophic," because all of the commands in a transaction "must be executed as if the commands were a single command."  However, these commands must still be executed sequentially, even if they won't be interrupted by other commands, and if the server crashes in the middle, some of these commands will have finished and others will not have finished, and the data will still be inconsistent.  How are transactions more useful, as Nutt claims, in that situation than simply executing commands as they arrive? (Heather)

19.1 OS Design Considerations

• Performance, trusted software, modularization, portability
• Pg. 773: Why is it so hard to write trusted assembly code that will execute faster than trusted C code? Aren't they just two ways of accomplishing the same thing, only we have more control when writing in assembly? If C is so much better, why have any of the kernel written in assembly at all? (Nick)
• When Nutt says that it is harder to write trusted code in assembly language that executes faster than C code, does he mean that what you need to do to prove assembly language is safe makes it run slower? (Lorelei)
• Pg. 774: When you compile or recompile your kernel, you can add support for something (say, a new filesystem) as a module or as part of the kernel. Is this module considered trusted software? If so, how is that security risk addressed? (Nick)
• [Portability] How does the source code organization in Linux allow it to avoid using indirect function calls? (pg 778) (Jonathan)

19.2 Monolithic Kernels

• Motivation?
• If the monolithic kernel approach works best with a small, simple trusted kernel, won't implementing OS functions outside this kernel require them to interact with it, creating indirection and performance costs? Is it just a matter of trading some performance for code that is more reliable? (Lorelei)

19.3 Modular Organization

• Would Object-oriented OS be very different to use(users)? (Sheng)
• Could you explain why many CS people hate Ms Windows in terms of its design aspects mentioned in the book? (Sheng)

19.4 Extensible Nucleus/Microkernel

•  Is the "extensible nucleus" organization of Kernel code just a combination of a base monolithic with modules added in? (pg 776) (Jonathan)
• I getting confused by the definitions of "operating system" when Nutt talks about the Mach OS. Specifically, Nutt uses this phrase to say "the Mach microkernel is intended to ... support different operating systems that have different policies" (786), but then goes on to describe key features of Mach as it it were an OS in itself. I always got the impression that OSes prized mechanism before policy, as the later can be changed by the user. How can these definitions be reconciled? (C.M.)

19.5 Layered OSes

• Does a layered architecture offer any opportunities to improve performance that the other types of architectures don't?  It seems like it would only result in increased overhead for almost every call and is mostly just a nice idea. (Heather)

19.6 Distributed OSes

• Does implementing a Distributed Operating System require that all the systems in the network be of the same type (e.g. all uniprocessors, multiprocessors)? It seems that the achieving transparency would be overly complex otherwise (p.798). (Angeline)

Janet Davis (davisjan@cs.grinnell.edu)