In attendance: Lingkun Chu, Aziz Gulbeden, Hong Tang, Tao Yang
(remote),
and
Jingyu Zhou.
Organizer: Hong Tang.
Time: 4:00-5:30PM.
The
meeting discusses the following topics:
(1) Project progress.
(2)
Evaluation benchmarks.
(3) Replication support.
(1)
Progress:
Aziz has almost finished
the
functionality of Server module (storage
management), and is refining
the
interface and perform stress testing. The
testing should be done
soon.
Hong has completed the name space server
client
module (part of the
proxy), the membership management module, he is
continuing working on the
integration of data location module into the
proxy.
Estimated time of
completion is one week.
In the
next
step, Aziz and Hong will work together on a object catalog
management
module
(which maintains a catalog of objects stored locally, will
be used for
data
migration), and a indirect location table (which stores
pointers to
objects
not stored locally). The object location scheme will
only use
consistent
hashing (which will lead to a host that either hosts the
object or
keeps a
pointer to the actual location of the object). This step
could take
about 1-2
weeks to finish (very rough estimation).
Hong will
also
work on the integration of Proxy and Server, and
implement an
adaptation
module for trace replay.
(2) Evaluation
benchmarks.
Jingyu will work on the NPB over MPICH
over
PVFS, and patch the
client-side module of PVFS to collect traces usable
for
evaluation. The same
tracing utility could also be used to get the I/O
trace
of other
applications running upon PVFS (maybe the BLAST package).
Jingyu
plan to
start work on it after he has finished his current work on
membership
management, which could take 1 to 2
weeks.
Lingkun will help on getting the I/O trace of the Retriever or
BLAST
running
on manually arranged partitions. Start time is not determined
yet.
Tao is proposing the Crawler/Indexer benchmark
(which
looks to me are
two benchmarks). There are interesting interactions of
these
two
applications (in some sense the provider-consumer relationship). He
is
going
to help on getting the statistics of the read/write rates of
these
benchmarks. We also need to think about how to devise some
generic
hints to
aid the intelligent placement of data (either way, the
intelligent
invocation sites of indexers).
(3) Replication
support.
Tao is also suggesting to have the
replication
support in the prototype.
Hong explained the replication and consistent
management schemes has already
been discussed in the design of Sorrento
(at
least in the tech-report). We
assume low write/write conflicts and use
version-based scheme to avoid high
overhead of synchronizing all
replicas.
Summary of the execution of other action items
listed in
the previous
meeting:
Aziz has expanded
the
plan document and maintains a project web site at
http://www.cs.ucsb.edu/~gulbeden/sorrento/INTERNAL/ . Be alerted, we may
implement some access-control
schemes
to prevent the place from being
visible to
non-members.
Jingyu has created a sorrento user group. Not clear about the status
of
the
sorrento email alias. Jingyu, can you report me on
this?
-Hong
Last modified: Fri Jun 20 00:29:41 PDT 2003