Research on OGSA-DAI/DQP
Part I
Alexander Wöhrer
U i
University
it off Vienna
Vi
Faculty of Computer Science
Institute for Scientific Computing
woehrer@par.univie.ac.at
Active vs passive DBMS usage
„
Passive DBMS usage:
„
„
„
⇒
Client
DBMS
DMBS acts as pure data
d
provider
id
all application code @ service level
pull mechanism for data to be processed
request
JDBC
result
Data
service
Active DBMS usage:
⇒
⇒
⇒
⇒
DBMS acts as technology and data provider
request
interface code @ service level
main application code @ DBMS level
Client
push mechanism for result data
OGSA-DQP Meeting,
NeSC March. 2009
Data Service
Interface
result
DBMS
Code
2
Active Database Integration:
g
Java in the Database
„
„
„
most of common vendors support Java Stored
Procedures in some way
Static methode of Java class as program entrance
point
P tG SQL 8
PostGreSQL
8.3
3
„
„
„
Oracle 10g
„
„
„
„
supports standard JVM
not much development support
comes with OracleJVM
implements a non
non-preemptive
preemptive threading model
rich support for Web-Service call-ins/call-outs via JPublisher
All have kind of server side JDBC support
OGSA-DQP Meeting,
NeSC March. 2009
3
Active Database Integration
g
Implementation
sendData(tuples)
Database host
Why tuple serialization?
„
First Oracle JDBC call
(evaluate with plan) blocks,
as at least one child thread
is still running
Evaluator WS
interface
serialize
tuples
Tuples in
Files
de-serialize
tuples
evaluate(query partition)
ExIn
DBMS
client
JDBC
SOAP
ExOut
„
Join
Table
OGSA-DQP Meeting,
NeSC March. 2009
server
JDBC
Scan
⇒
PostGreSQL wouldn‘t
wouldn t block
but we had to block it,
otherwise „forgotten
statement“
these restriction lead to
tuple
p serialization via files
for both implementations
4
Lessons learned…
„
OGSA-DQP 3.2 TP tends to over-parallelize
„
„
„
due to poor estimates
d
i
off static
i distribution
di ib i costs
average row size provides insufficient information
average column size needed
„
„
„
providing the base information to calculate better
distribution costs
memory requirements
avoid service hops
„
Getting rid of artifical OGSA-DQP evaluator layer
on top of OGSA-DAI results in
OGSA-DQP Meeting,
NeSC March. 2009
5
Austrian Grid
„
Medical application
pp
„
„
breath gas analysis
Focus on security issues:
„
„
Single Point of Access (SPOA) cloud
computing concept based on OGSA-DAI
and Nimbus
Fine-grained data access for row level
security
OGSA-DQP Meeting,
NeSC March. 2009
6
China‘s Railwayy Freight
g
Transportation Grid
„
„
One year visitor PhD student Yan Zhang @ ISC
Application characteristics:
„
„
„
„
„
dedicated network
data integration and
analysis
1 waybill record
around 4kb
150.000 per day
220 GB / year
OGSA-DQP Meeting,
NeSC March. 2009
7