5/30 paper evals:

Yod (h13nguye@ieng9.ucsd.edu)
Tue, 30 May 2000 02:23:38 -0700 (PDT)

Henry H. Nguyen
BS Computer Engineer
ME Computer Engineer
(858) 587 - 7046
Title: Implementing Global Memory Management in a Workstation Cluster

This paper attempts, also claims, to implement the optimal approach to memory
management by integrating algorithms at the lowest level of the system to
encompass all memory activity such as VM paging, mapped files, and explicit
file access.

The algorithm in this paper uses the concept of local pages and global pages to
distinguish between pages on a node in the system. Pages may also be private
or shared. The algorithm changes the local/global memory balance as the
result of faults caused be an access to a nonresident page. And the cost of a
memory reference depends on the state of the referenced page such as local,
global, or on disk. A local hit is over three orders of magnitude that of a
global hit, which is more than ten times faster than a disk access.

Page replacement is simple, faulted page is read from disk, and the oldest page
in the network is either discarded or written back to disk. The oldest page is
determined by an initiator node which maintains the age information of every
node for both local and global pages.

The paper goes on to describe the implementation of the system. It talks about
the three basic data structures of the system: page-frame-directory, global-
cache-directory, and page-ownership-directory. Inter-node communication is
through the assumption that the network is reliable, which enable marshaling and
unmarshaling to and from IP datagrams directly.

The last issues mention in this paper are performance and limitation of the
system. Tests were set up to measure the performance of the basic instructions
of the system such as getpage and putpage operations. And more comparisons were
made between the system with GMS and one without GMS, or GMS versus NFS. The
limitations of this system includes: have not yet handle the failure of the
initiator or master nodes, and the blind trust that one node has for another to
not reveal or corrupt its data that is stored in the second node's global mem.

Title: Memory Coherence in Shared Virtual Memory Systems

The topic of this paper is memory coherence in shared virtual memory systems.

Shared virtual memory, as describes in the paper, is a single address space
that is shared among all processors in a loosely coupled distributed-memory
multiprocessor system. Any processor can access any memory location in the
address space directly. Memory mapping managers implement the mapping between
local memories and the shared virtual memory address space.

And the two issues listed in the paper for the implementation of a shared
virtual memory are: the granularity of the memory units and the strategy for
maintaining coherence. The granularity of the memory units should be consistent
with that used in conventional virtual memory implementations to take advantage
of the existing page fault schemes. However, the larger the memory unit, the
greater the chance for contention.

Memory coherence strategies, the main focus of this paper, deal with the issues
involving page synchronization and page ownership. The two basic approaches to
page synchronization are: invalidation and write-broadcast. The different btwn
these approaches is that there is only one owner processor for each page, but
a page can have several owners in the write-broadcast approach.

Page ownership is decribed in detail in this paper. Page ownership can be fixed
or dynamic, and dynamic page can be centralized or distributed, and distributed
managers can be fixed or dynamic. Then there is the use of Info table and
PTable to keep information about the accessibility of pages. The paper goes on
to elaborate on the implementations of the various page ownership methods,
including the use of hashing function to determine which pages are mapped to
certain processor in the distributed manager method.

The paper also includes tables from the experiments that were designed to sup-
port the memory coherence algorithms proposed in this paper. Their tests showed
that dynamic distributed manager algorithm and its variations seem to have the
most desirable performance.