SUNY at Stony Brook, Computer Science Department
Eugene W. Stark,
$7,450 on 9/26/97
To purchase equipment to set up an experimental platform to do development
and evaluation of a transactional shared memory facility.
Transactional Distributed Shared Memory Abstract
Transactional distributed shared memory (TDSM) integrates transactions
and distributed shared memory, with the purpose of endowing DSM with fault
tolerance attributes. The main goal of our research is to design and implement
a highly available TDSM facility, so that it can be used to build distributed
applications that need to share somewhat large amount of data, e.g. a name
service, in a network of workstations environment.
Current Status as of 3/98
We have implemented a prototype of the TDSM facility for the FreeBSD
2.1.0 operating system and we are finalizing its port to FreeBSD 2.2.5.
This prototype provides "distributed" nested transactions management and
distributed lock management. However, it is not fault tolerant to either
host failures or network partitions. For example, failure of a host may
lead either to the loss of data that was modified by committed transactions
or to the unavailability of data on which that host holds locks. We have
designed a protocol to provide durability via data replication in different
servers upon transaction commitment, and will start its implementation
as soon as we are done with the porting to FreeBSD 2.2.5. We have also
started to address the availability issues, by considering some possible
algorithms to recover locks held by hosts that fail or that become unreachable.
But we need to work further on this latter problem.