Extended Discussion

Freeman Rawson discussed the failure of Workplace OS, by his account a huge project at IBM. Its ambitious goal was to replace all IBM OS'es in the 1990s (OS/400, OS/2, AIX). It would be scalable from the palmtop to the minicomputer and portable across platform and process. It would be based on a microkernel, make extensie use of OO, and be marketable in seperate parts (e.g. microkernel seperate than OS). Freeman emphasized that this was a development project and not research and that they actually created a shipping version before the plug was pulled on the project. The numbers that came out from the performance guys were not encouraging and the measurement were done on commodity software.

The Workplace OS was based on the Mach 3.0 microkernel using Julin's multi-server design. They ran multiple personality servers on top of it so that they could provide compatability with multiple OSes. To support this, they needed to factor semantics and function from the different OSes into each server. For example, the file system server had to support both AIX, OS/2 and OS/400 semantics. One of the key observations made was that it's hard to factor operating system semantics and that it's not reasonable to support all semantics of an OS in a small space. Another observation is that compatability is in the details, and the details take a lot of time and effort.

Rawson went on to argue that hardware architecture conflicts with microkernels because of the cost of the address space switches and despite all of the optimizations they attempted, this was still a major overhead when running commodity applications on top of Workplace OSes.

Rawson ended up giving some advice to the research community. He suggested the community spends too much time worrying about Distributed Shared Memory (DSM) and large-scale parallel systems. He suggested that device drivers, being a painful problem for the OS implementor, were worthy of much more research work.

His parting wisdom on the Workplace OS project was that development "should never mistake reserach projects for products." As for the use of Mach in Workplace OS, he pointed out that code reuse is great but design re-use is better. Neither is everything in end.

Discussion

Brett Fleish also gave a talk about Workplace OS. He sprinkled his talk with general insights about engineering and failure with many examples taken from the pages of Henry Petroski. His main argument, though, was that personalities had somehow failed to generalize in Workplace OS. However, according to analyses, these personalities turned out to be impossible to structure in a way that maintained performance, because of the need to accomodate semantics from all the operating systems. He ended his talk poiting out that failure is a great mechanism for learning and that Workplace OS was arguable a joint academic/industry failure. He admonished the industry for not studying failure more.

Discussion

• Brian Bershad agreed with Rawson's slide on research areas. He pointed out that his group had submitted several papers on the areas of device drivers and emulation and that they had been summarily rejected. As such, there was a great disincentive to study this. He also took issue with the statement that it was academia's failure. He said papers had been published warning not to pursue such a design.

• Rawson responded by saying that he hadn't read such literature. He emphasized again that the need to lose the state of the processor while transitioning address spaces was a major source of pain.

• Liedtke pointed out that the cache doesn't necessarily get flushed on context switches.

• Rawson said the original view was that the PowerPC would be fast enough so the overhead wouldn't be a problem.

• Chris Small asked what would change now in 1997, given that the dominant personality is Windows, what areas would you work on? Rawson answered that more uniform and general-purpose resource management, real-time, and embedded systems were big areas. He also said that there would always continue to be research on non-Windows OSes.

• David Cohn asked when IBM knew that personalities wouldn't work? Rawson said it only became apparent in the last year of the project.

• David Presotto remarked that it's hard to turn a charging elephant.

• David Black took exception with the issue that microkernels couldn't be used efficiently. He pointed to joint work between the OSF and the U. of Utah. He also took exxception at Golub's paper. He said it compared apples and oranges - an early version of SunOS 4 versus the BSD server on top of Mach. Sun had already made extensive changed to the BSD internals. Also, the system path length studies factored boundary crossing too heavily in the system call.

• Presotto - flogging a dead horse if I've ever heard one.

• David Black pointed out some shortcuts they did to make OSF preform. First of all, they no longer used an emulator, since that made signals hard to implement. Co-locating the BSD server with the application shortened the system call path.