************************************************************************
UNIX as an application program
************************************************************************

This paper describes Mach operating system. Mach is a micro-kernel. This
means that the kernel implements only the very basic services needed by
user applications, and all higher-level operating system services, such as
memory management, file systems and pagers are implemented above the
micro-kernel in application servers. 

The motivating factors behind micro-kernels are flexibility, extensibility
and portability. Mach, with a UNIX server running on top is structured as
follows: the Mach kernel provides IPC services, memory object creation,
scheduling, trap redirection and low-level device support. All other
functions usually associated with the operating system are provided by a
user-level process - a Unix server.  The applications running on Mach have
to use a Transparent System Call Emulation library. This library runs in
the address space of the application and redirects all system calls made
to the operating system to the Unix server via Mach IPC facility. The
emulation library code is shared between parent and children
processes. Requests for service come to the Unix server in the form of
Mach messages that request a particular Unix operation. A thread is
dispatched to handle each request. 

The implementation described in the paper uses a very interesting
optimization for handling I/O. After doing many tests, the authors
realized that copying data on I/O was prohibitively expensive. They
decided that they should map all files directly into the user address
space, which is managed by the Transparent Emulation library. This enabled
most all file system calls to be executed right in the library without
ever contacting the server.

The experimental section of this paper is pretty weak. Most tests compare
the previous release of Mach with Mach running with Unix server. There are
almost no comparisons with other top-of-the-line operating systems. There
is one test, compile test, that compares performance of Mach with SunOS
4.0, however the results are not explained at all.  Most of the tests
presented in the paper involve file system operations - and this is the
piece of the system that has been heavily optimized. It is not fair,
therefore, to test only the most optimized part of the system. It is clear
that the reader would want to know how other types of applications
perform. For example, the table comparing process management costs under
Mach and SunOS 4.0 shows that getpid and fork/exit operations are more
expensive under Machs.  I felt that the authors did not study the
advantages and disadvantages of micro-kernel thoroughly: they did not show
when one can gain and lose in performance, and they did not provide
motivation for using micro-kernel when there is performance penalty
involved.