cse221: paper evals

Octavian Luca (oluca@cs.ucsd.edu)
Tue, 30 May 2000 04:28:41 -0700 (PDT)

Multics VM

The authors of this paper set out to discuss the properties of an
"idealized" Multics memory cromprised entirely of segments referenced by
symbolic names. They describe how such an idealized system could be
simulated using specialized hardware and software. This work differs from
previous efforts in it's aim to achieve the effect of large segmented main
memory and direct addressability making information copying no longer
mandatory. Thus as opposed to other systems, access control to shared
core images is no longer necessary, and therefore it no longer requires
additional hardware.

The paper first describes the two goals of the system: making it possible
for all on-line information stored in the system to be directly accessible
by the processor and controlling access, at each reference, to all on-line
information in the system. Then the path taken to achieve these goals is
described in greater depth by first explaining the motivation for using
segmentation as a facilitator for sharing information without duplication.
Then the paging mechanism is explained and the role of supervisor
processes in VM management. The tables used store the information for the
proposed implementation are described and how the hardware complements
wherever possible their structure.

Of note this system's segment attributes are stored in directories
organized in trees whose structure is known to the user and accessible
through symbolic names. Also the large segmented memory is directly
addressable through symbolic names and therefore any access to memory is
mitigated through the access checing mechanism. This paper was of
interest in the way it took an exisitng system and through some
modifications achieved some desirable design goals. It seems that further
work in this area is possible, with the goal of improving the efficiency
and performance of VM through insights into segmentation.

Global Memory Management in A Workstation Cluster

This paper describes the design and implementation of global memory
management in a workstation cluster with the aim of using a distributed
management algorithm at the lowest levels of the os. The researchers
believe that such a system has many advantages over conventional
implementations where there is no cooperation among machines. One such
advantage is that global memory management avoids unnecessary disk
accesses which is believed to become increasingly important with the
widening gap between processor and disk performance.

The authors list four important differences between their work and
previous efforts: their use of the algorithm at the lowest system level
and encompassing all memory activity, their use of globally controlled
resources versus independent workstations, addition and deletion of nodes
without user intervention, and implementation well integrated into an
existing OS.

The paper first describes in detail the algorithm used to implement shared
memory and explains various important aspects of the implementation. Then
the performance of the system is measured using a series of
microbenchmarks focusing on bookkeeping overheads, execution time
improvement, responsiveness to load changes, etc. Some system limitations
are also discussed such as the impact of node failures on the overall
system and the negative impact of the global memory management system on
applications running on non-idle nodes.

This paper is yet another example of research taking advantage of progress
being made on the hardware front. In this case faster networks allow
efficiency improvements in the face of the widening gap between cpu and
disk performance. This topic seems to be of increasing importance as the
trends discussed in this paper do not seem to reverse any time soon.