Paper Evaluations

Carnevali $ilvio (csmede@hotmail.com)
Wed, 03 May 2000 19:05:52 PDT

Geoff,
The Sprite paper mentions that LOCUS, ACCENT and V all implement process
migration (page 26, 1st column, 3rd paragraph), but I did not find such
evidence while reading the respective papers.
Furthermore, last Tuesday you told me that processes can be created remotely
in LOCUS but cannot migrate, which I think is true after re-reading the
concerning paragraphs.
Any answer to that dilemma? Is it possible that the LOCUS paper implements
process migration even though it does not mention it?

And here are the evaluations:

V Kernel

This paper discusses the main features of the V Kernel, a message-oriented
OS supporting local and network IPC.
The interesting idea it introduces is the concept of diskless workstations,
based upon network file access, which
is implemented in a layer separated from IPC; this is one of the main
differences with LOCUS.

The V Kernel provides a variety of primitives for interprocess
communication, specifically designed to optimize
transfers of small as well as large portions of data. Messages are
fixed-lenght (32 bytes), but transfers of
large files can be accomplished using specific functions (Moves).
Synchronization of communication is also
optimized as the ACK signal and data can be sent together; thus, only one
ACK is needed when sending large chunks
of data.
The use of a global naming strategy for processes is useful for determining
whether a message was sent by a local
or remote process. In the latter case, efficient local operation can be
achieved, which is one of the goals of the
system.
The authors focused on the V Kernel statistics in an effort to show that
local performance was comparable to that
of other existing systems; remote performance was also claimed to be
dependent upon network penalty, and thus the
conclusion that the V Kernel was competitive wrto other systems was reached.
It is interesting to note how the authors have tried to use arguments like
parallel computing during network
communication as well as high disk access times to show that the V Kernel's
remote overhead was mainly due to
parameters independent of the implementation of the system.

This paper helped me gain a broader view of what is implied in the design of
a message-oriented system, in comparison
to other systems we've been studying until now. I think it gives a good
explanation of the basic functionality of
the system as well as accurate performance measurments; this is probably the
reason why there is no mention of future
work, since the main focus is on the actual state of the system.

7'UP

Sprite is "yet another network OS" (yanos - in contrast with yacc). The main
goal of this system is taking
advantage of the recent technological advances concerning network, CPU
performances and memory size in order
to hide (reduce) the overhead for remote access. Probably the most
innovative aspect of Sprite is the
efficient implementation of process mobility.

Even though the authors claim that the system was designed from scratch,
there are many features that were
imported directly from UNIX. Other features were added to the system though,
like data, physical memory
and CPU sharing, process migration, as well as complete file system
transparency similar to the concept
used in LOCUS.
For data sharing, processes use either a UNIX-like fork operation for code
sharing or the new concept of shared
fork when code and data sharing are neded.
Physical memory management introduces a new page replacement policy: file
caching is separated from virtual
memory, but they are closely related since each module can replace the LRU
block of the other module.
Process migration is an interesting concept that increases overall system
performance as it enables a better
CPU sharing policy: processes on overloaded workstations can migrate to an
idle one with little penalty;
communication between them is performed via a special RPC facility that uses
Client Stubs and RPC transport units,
thus making communication completely transparent to the moving process.
Finally, file transparency is achieved using prefix tables, a generalization
of the UNIX static concept of mount
tables. The main advantage is that any domain changes are dynamically
updated in the prefix tables.

This paper presents clearly the main OS architecture trends used at that
time. Many of these concepts are still in
use today; I personally learned a lot from this paper and can affirm that
many of these concepts are clearer in my
mind now.
The authors mentioned that the System was almost fully implemented, even
though some details about process migration
were not fully functional yet.

________________________________________________________________________
Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com