Ev

qin Zhanhai (qinzhan_hai@hotmail.com)
Tue, 30 May 2000 15:31:45 GMT

--------------------------------------------------------------
Memory Coherence in Shared Virtual Memory Systems
--------------------------------------------------------------
The paper discussed the well-known memory coherence problem on
a loosely coupled multiprocessor. And by proposing two classes
of solutions and the corresponding experiment results, they
justified that the problem can be solved efficiently.

"A memory is coherent if the value returned by a read operation
is always the same as the value written by the most recent write
operation to the same address." They proposed two classes of
of solutions to memory coherence problem: centralized memory
mapping manager algorithms, and distributed memory mapping
manager algorithms. In the former algorithms, mapping manager
was put on one processor. In other words, there is only one
manager that knows who the owner of a page is. As for distributed
manager algorithms, the mapping managers were placed on all the
processors. The owner could be found by fixed mapping(Hash function)
or dynamic mapping(ProbOwner field).

-------------------------------------------------------------
Implementing Global Memory Management in a Workstation Cluster
-------------------------------------------------------------
The paper proposed an single, unified, but distributed algorithm
of global memory management in a workstation cluster. The global
memory management algorithm was inserted at the lowest OS level.
The system can automatically reconfigure to allow machines to
join and in most cases, to depart the cluster at any time.

The pages on a node were classified as being either local pages,
which have been recently accessed on the node, or global pages,
which are stored in the node's memory on behalf of other nodes.
Pages may also be provate or shared. The algorithm handled four
cases by changing the local/global memory balance as the result
of faults caused by an access to a non-locally-resident page:
1) The faulted page is in the global memroy of another node;
2) The faulted page is in the global memory of node Q, but P's
memory contains only local pages;
3) The page is on disk;
4) The faulted page is a shared page in the local memory of another
node.

The page replacement algorithm was probabilistic in a sense that
global age information of a page was maintained probablistically.

One characteristic of the algorithm was that node failures in
the cluster do not cause data loss in global memory because all
pages sent to global memory are clean.

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