CSE 221: Paper Evaluations

Renata C TEIXEIRA (teixeira@cs.ucsd.edu)
Mon, 29 May 2000 20:48:52 -0700 (PDT)

Title: "Memory Coherence in Shared Virtual Memory Systems"

This paper proposes algorithms to implement shared virtual memory
on loosely coupled multiprocessors. The main focus of the paper
is on memory coherence.

There are different strategies to maintain coherence, the ones
discussed in the paper use invalidation approach for page
synchronization and dynamic approach for page ownership strategy.
A mapping manager implements the mapping between local memories
and the shared virtual memory address space. The strategies for
maintaining dynamic page ownership can be subdivided into two
classes: with centralized or distributed manager. The distributed
strategy can be fixed or dynamic. The algorithms use page fault
handlers, their servers, and page tables.

The paper presents a thorough description of the algorithms,
proposing improvements to them, proving their correctness,
and implementing a prototype. The idea of using a distributed
approach is appropriate for systems with a large number of
processors because it scales better. The paper does not
discuss, however, how failures and message losses are handled
on the algorithms.

Title: "Implementing Global Memory Management in a Workstation Cluster"

This paper describes the design and implementation of global memory
management in a workstation cluster. The idea is to manage clusters
more as a multicomputer than as a collection of independent machines.
The algorithm is integrated with the lowest level of the operating
system and encompass all memory activity: VM paging, mapped files, and
explicit file accesses. Besides, the algorithm can gracefully handle
addition and deletion of nodes in a cluster.

Pages on a node P can be either local, which have been recently
accessed on P, or global, which are stored in P's memory on
behalf of other nodes. Pages may also be private or shared.
The algorithm uses a replacement scheme based on cluster-wide
age information. Thus, idle nodes will store more global pages,
working as cache for more active nodes. This mechanism avoids
fetching pages from disk and hence minimizes the total cost
of all memory references.

The paper discusses in detail the performance of the system and
compare it with another system. From this evaluations I could
see the benefits of the algorithm proposed, which was not so
obvious by the description of the algorithm. One interesting
idea is the cluster-wide age information distributed periodically.
This information allow a good replacing strategy with less
overhead on the system.

Renata Teixeira