[Pgi-wg] Notes of OGSA_BES session at OGF27
Johannes Watzl
watzl at nm.ifi.lmu.de
Fri Dec 4 09:30:30 CST 2009
Dear all,
please see below the notes from the BES session at the OGF27 in Banff.
Best,
Johannes
Notes:
Requirements Cream/ARC
Steven:
GLUE2: XML rendering not ready
not talking about attributes
Andrew:
just relabelling
Morris:
in description of epr
problem in GLUE2:
circle dependecies
an epr can have numerous resources and one resource can have numerous ep
Andrew:
in originalBES spec: no info
Donal:
retain info for how long?
Andrew:
til it is finished
not make a minimum - imagine 10000 jobs -> huge amount of data
you should not be able to see other people's eprs
PGI requirements
createActivity
pending -> running
stay there?
was put out the original
pending, running, failed
list, delete, copy, write
Morris:
refine -> more detail in fault
Philipp:
lot of checking needed
Steven:
if element not supported -> throw fault
now: bit more information
Andreas:
no endpoint in JSDL
Andrew:
enumerate classes of problems
Morris:
do we want them in the spec?
Steven:
session directory
Andrew:
what is session directory?
Steven:
URI is returned to client
see from state machine -> kick off exec
Andrew:
rather profile
Morris:
never using service staging in
Steve:
avoid roundtrip
David:
no staging in
just running?
Andrew:
doc that comes in not for JSDL
-> activity doc that includes JSDL
Philipp:
activity ep doc
Andres:
BES ep profile
Morris:
when decide start/run/stop
hold point
spec. JSDL
S0 -> S0' -> S1
stop here
Andreas:
data staging described outside
Morris:
should it be interface or JSDL issue?
Andreas:
complex workflow
not good to put it in JSDL
Andrew:
stop:hold points
Morris:
also pre- and post processing
- call preproc, put data right
- job run
- post proc
Andrew:
not putting complete workflows
Andreas:
not in JSDL
-> specify
Morris:
working dir in JSDL but not working
you don't know where job is running
Dave:
underlying implementation
Andrew:
want run workflow on node!?
Morris:
single one
Andrew:
solve problem:
you need to come back to same container
Morris:
one BES; 4,5 nodes
Andrew:
local data management
how people mount file system
Steve's suggestion:
BES container epr
Philip:inUNICORE after job is done all dirs are cleaned
more like an implementation issue
session dir is no problem
Steve:
JSDL
control: do it or don't do it
finer control wanted
Andrew:
data motion is also a job
determine, add requirements
how are we going to handle these problems?
look at different mechanisms
have working dir, can persist
potentially run jobs in session dirs
-> data close
external reference
purge, create seperately
give scratch space
Morris:
what if lots of CPUs needed later?
Andrew:
scheduling decision
David:
two stages
Andrew:
two function calls: slower
Morris:
epr must be refined
Steve:
do you have a simple BES ep?
represent resource sharing in BES interface
Andrew:
BES has management interface
Steve:
Why does share of resources need to be reflected?
Andrew:
Behind the scenes we can have different BES for different queues.
-> separate BES resource for each phys. queue
Morris:
In EGEE: one BES container somewhere
David:
queueing on VOMS attributes
job manager does automatic queuing
WMS does not include queue name in the job description
Steven:
Why not reflected job requirements in JSDL doc?
Morris:
maybe
Andreas:
feature where are you submitting to
David:
complete different ways in lsf and Condor
Andrew:
that's why different BES
Steven:
BES endpoint
multiple resources
Morris:
JSDL: specify share
keeping contacts only once
Andrew:
encapsulate or not?
Steven:
GLUE2: coordinates of info about resources
-> GLUE2 service ep
BES ep described through GLUE2
Morris:
different containers
Steven:
different endpoints
Donal:
WS-addressing
Morris:
WSRF
Andrew:
WSRF and WS-addressing are not coupled
Morris:
reference parameters
black box to users
Andrew:
two choices
* let PGI dow what they need for the specification
* BES 1.1
going to BES 1.1
coupled with JSDL
Morris:
slide "Refinement concepts overview"
David:
information service
Steven:
What to do with GLUE2?
Steven:
job info in GLUE2 scheme
Andrew:
we talk to endpoint to find out the requirements
push or pull model?
for dynamic meta data push model
good notification system
Andrew:
HPC resource?
not an OGSA-BES thing
Steven:
back to session dir
Andrew:
one execution line to multiple execution lines
Andreas:
it is a workflow
Andrew:
you can do this with BES: Data-Transfer delegation
when BES service is running it does delegation
Philip:
activity instance schema
keep history what happened to activity
Andrew:
send log info to some place
Philip:
use case from D-Grid
Platform
UNICORE
Andrew:
profile
GLUE2 used for?
Andrew:
GLUE2 schema not an implementation
Alexander:
if you submit activity doc you get back an endpoitn
Steve:
will there be something from BES 1.0 in BES 1.1
Andrew:
everything will change
comepletely new namespace
supporting both port types in the same service
Steven:
backward compatiblilty
implementing the two different port types
Andrew:
1.1: superset of 1.0
implement 1.1 machinery on 1.0
Steven:
client with 1.0 should be able to use 1.1
next steps:
- Group is revising
- Strawman within BES
- Work on the concepts one at a time
- Not starting with delegation
- Different implementations
--
_ _ _ _ _ _ Johannes Watzl
|\/| |\ | |\/| Institut für Informatik / Dept. of CS
| | | \| | | Ludwig-Maximilians-Universität München
======= TEAM ======= Oettingenstr. 67, 80538 Munich, Germany
Room D0.5, Phone +49-89-2180-9162
Munich Network Management Team Email: watzl at nm.ifi.lmu.de
Münchner Netz-Management Team http://www.nm.ifi.lmu.de/~watzl
More information about the Pgi-wg
mailing list