IPSJ 72nd National Convention, Day 3

Today, I attended at a regular session at the IPSJ national convention.

https://www.gakkai-web.net/gakkai/ipsj/72program/html/program/P5.html#6E

Here are my notes.

6E-1
Hum... Their work looked like at very early stage. I'm not
sure why they did the presentation now...

6E-2
It looks like a part of METI funded national R&D project carried on
by NEC Laboratory. The work is about a new (?) cache data control method
for resource configuration information expressed in CIM. It's a variation
of LRU resource management techniques, I think, and their basic idea
is to use residual time before expiration of various resource
configuration information. The experimental results show the validity
of the idea, but I felt it's important how to choose the TTL values of
each resource configuration information. I pointed out it after their
presentation, but

6E-3
A work for realizing multi-tenant environment without
any modifications of user Servlet applications
done by Hitachi System Development Laboratory.
The essential point of the work is that keep using a single database. :o
I would have separate databases for each tenant and just modify DBCP
(database connection pool) configuration, but it looks like
the strong constraint they assumed came from GlassFish.
Technically, their implementation stores tenant level context information
in thread local storage, intercepts some SQL statements and rewrite
table names in the single database.
Performance impact seemed to be small, but it's better to make longer
pseudo application loads than they did.

6E-4 Canceled.

6E-5
Our work on cloud federation manager which I mentioned here several times.
Actually, what we thought was that we put too much information
into our paper. :o

6E-6
An interesting work on multimedia processing system by
NTT Cyber Space Laboratory. They have been working
on a huge scale stream data processing cluster engine called DMS for
several years, and this work "Niji-Gumo" (虹雲) which means rainbow
cloud is an implementation of a new frame work supporting intuitive UI
framework using DMS as its backend engine.

The work is interesting, but it was not clear for me what theme the presenter
wants to focus on. User interface improvement, multimedia data processing or
DMS itself...?

6E-7
A work on improving a multi-tier streaming data processing engine, DMS,
from the same organization with the previous session.
DMS is basically consists of 3-tier clusters, and the second tier which
is responsible for the most CPU intensitive processing *pulls* requests
from the front tier. Furthermore, members of the second tier do not
communicate with each. This structure comes from a motivation shooting for
maximum parallelism of course. But, it naturally produces a side effect
that sometimes a lot of the second tier servers try to pull requests
from one front end server which the called 'back-pressure problem'.
This time, they chose to solve the problem making a kind of waiting queue
in the front end tier servers. :(

I thought the resolution is quite ad-hoc and they should consider the puller
side resolution too because nowadays there are good algorithms for keeping
consistency in distributed environment.