Present: Paolo Capiluppi (INFN/Rome),
Mauro Campanella (INFN/Milano),
Dirk Duellmann (CERN/RD45), Philippe Galvez (Caltech),
Irwin Gaines* (FNAL), Frank Harris* (Oxford), Ian McArthur*(Oxford),
Harvey Newman (Caltech), Bianca Osculati* (INFN/Genoa),
Les Robertson (CERN/IT), Simona Rolli (FNAL),
R D Schaffer (CERN/RD45), Leonello Servoli (INFN/Perugia)
Jamie Shiers (CERN/RD45), Krzysztof Sliwa* (Tufts),
Andrei Tsaregorodotsev (Marseilles),
Richard Wilkinson*(Caltech), David Williams (CERN/IT)
[*] By video
Introductory
Note: These minutes are extracted from a long and somewhat
involved
meeting. MONARC members and especially the Insitute group
leaders are requested to read these minutes carefully, and
o H. Newman opened the meeting noting that the conclusions of
the meeting will be
preliminary in the sense that Laura Perini
could not be present.
An initial go-round of remote participants
indicated that there were various audio
problems, some of which may have been associated
with the reconfiguration of the
transatlantic network and the fact that there
were 2 other videoconferences going on.
Action Item:
In all future meetings, technical tests and setups will commence
30 minutes before the start of the meeting. It is not advised to try
to just join the meeting at the moment the meeting starts, especially
if the user and/or the setup is "new".
The third MONARC meeting, shortly before the
release of Draft -1 of the PEP is
scheduled for
Tuesday August 18 at CERN at 16:00, with Virtual Room Video available.
It is proposed that Laura Perini chair this meeting.
The MONARC name was accepted
o Before the meeting Mauro Campanella submitted some of his ideas,
distributed just
before the meeting, derived from his note
"The analysis model and the optimization
of geographical distribution of computing
resources: a strong connection" by him and Laura, which may be found at
http://www.mi.infn.it/~cmp/modelsv1.html
Mauro noted that many Models can be ruled
out prior to any detailed simulations,
and emphasized the need for more "paper
Models" of the Analysis Process. A Model
often requires too many resources in
at least one area, or turnaround times are too
long if one presumes to do certain operations
(example: too-frequent re-reconstruction)
during the course of the analysis. He
then noted the
advantage of information from currently
or soon running experiments in choosing a
sensible initial guess (or guesses)
at the Models to be simulated, and suggested that we
explore finding collaborators from these
experiments. [I note that several can be found
amongst those on the MONARC mailing
list.] Frank Harris expressed a similar opinion
after the meeting, by E-mail.
Set up team to begin looking at Models on paper, and eliminate
those that are clearly unreasonable.
In summary, he emphasized the need to
focus on a few areas in the Working Groups:
- choice and
validation of at least two simulation systems;
- analysis models
input parameters (experiments wishes/goals, data
on present experiments);
- analysis of
possible models;
- database performance
on WAN, setting up a distributed testbed;
o Harvey presented a proposed PEP Outline, and a
detailed presentation of an
idealized set of Phases, Tasks
and Subtasks for the project. His slides may
be found at
or as a Powerpoint file as monarc730.ppt in the/july3098/ directory above.
The Phases were foreseen to follow
a linear plan including Setup, Startup, Modeling,
Refined Modeling, and Verification
and Convergence, all between 10/98 and 12/99
when the Computing Technical
Progress Report was foreseen to be submitted.
There are many items to setup,
parametrize, install, manage and bring into operation.
The behavior of the Object
Database and HPSS are complex and will be hard to
abstract. It will be hard
to focus on sensible Models early, and there may also be
several "branch points"
where Models may be inherently different (example:
centralized versus highly
distributed Models) and where the analysis of different
Classes of Models may persist;
perhaps for most of the Project.
Paolo commented that the plan as
shown was quite "linear" compared to the reality
where the study goes through a
development cycle (Simulate/analyze/evaluate/refine/
resimulate, etc.). This is clearly
true, but it was agreed that a defined project has to
assume that each of these Study+development
loops last for only a limited time.
He also expressed concern, supported
by many, of the tightness of the schedule given
that there were many geographically
dispersed people, nearly all with a limited time to
work on the project. In this light
the limit on the Project to 12/99 could be unrealistic.
A review of the Tasks in the plan
presented by Harvey led to more general concerning
about the depth of the problem(s)
to be solved and the limited time available. Dirk
asked about how the studies might
account for certain major "events" such as loss
of contact with a database partition
over an extended period. It was said that the Models
would have to assume that the
major components, including the database, had to be
reliable. RD45 was requested to
provide information as to the "extra" resources
required to robustly mirror the
data, to satisfy the (heavy) requirements for reliability.
Apart from the specifics, this
discussion served to illustrate that even "refined"
Model studies will fall short
of the real level of complexity of the distributed
systems the LHC experiments will
build.
Mauro emphasized that there will
be intense activity in the early days of LHC
running, until the analysis and
software are in a somewhat stable "production"
status. One cannot assume that
the lower data rate will mean relatively low
demands on the Computing Model
in 2005 compared to when the LHC is running
at full luminosity. This comment
served to illuistrate that certain uses of the
resources will be restricted,
and strict priorities might have to be set by each
experiment as a matter of policy
(example: the Production Reconstruction group
gets absolute priority if resources
are not adequate to cover all activities in the
data analysis). Such policy decisions
should be reflected in the Models where
relevant.
As a result of the discussion it
was agreed that MONARC would not start out
by requesting delays from the
LCB, but the scope of work would have to be moderated
to be more realistic. Since the
primary goals of MONARC are to identify and
classify some (not the unique)
"plausible" Models, the nature of the milestones
to be achieved within the allotted
time of 12-14 months could be more flexible.
In fact, the Computing Model considerations
reported in the CTPRs at the end of 1999
could simply reflect the state
of advance of the MONARC project. The issue of
whether the CTPRs should be a
few months later it order to take advantage of
a more adequate set of Model studies
by MONARC will be for the LCB to decide
during 1999.
o Harvey presented a first proposal for Working Groups,
as the basis for discussion
(see his slides for further details
on the tasks of each WG):
Systems Design, Analysis and Network
Process Design, Simulations, Model Measurements,
and Steering
David Williams noted that Systems
(or Site) "Design" might be better-called
Systems Parametrization; and similarly
for the Analysis Process. It was noted that
"Design" in the context of a Simulation
study of a complex system with different
levels of granularity, could mean
as little as the specification of a few of the main
components, their performances
and interconnections (e.g. processing power, data storage
and I/O channel speed for a computing
system).
There was some division of opinion
as to whether there should be just three, larger WG
called Systems, Processes and
Steering or whether the proposed structure should be maintained.
The reason for relatively small
WG was that people can work and intercommunicate better
in a small group dealing with
a limited number of issues. Sufficient frequent meetings would
be the way in which adequate inter-WG
communication would be assured. There were different
views expressed on the focus of
the Project: to get down to the simulations of Models, even
if the assumptions in the first
ones were not so realistic (I. Gaines, C. von Praun); or
alternatively to spend focus on
getting a reasonable starting point by paper studies of the
Analysis Process, and by detailed
discussions of the possible site-architectures and funding
constraints foreseen for 2005.
The Steering Group, which everyone agreed on was needed,
would have to settle on the focus
of the MONARC project plan and schedule, preferably
during the month of August.
From a practical standpoint it
became evident that the structure and course of the
project would depend greatly on
the manpower actually available, and the interests
and resources of each group in
MONARC.
Each group leader in the Project is requested to formulate and submit
a response from his Institute to the MONARC management by August 17
at the latest.
It was noted that the
Structure of the project need not be the same as a function
of time. Once a focused
series of simulation studies was underway, it might be
advantageous to reorient
the WGs and the balance between them.
A tie between these
considerations and the need for an officially-adopted CERN
vision of LHC Computing,
including a clear view of its scale and unique character
relative to nearer-term
HEP experiments, was noted.
The relationship of
the MONARC to CERN/IT was discussed. There are
a number of areas
where direct involvement by CERN/IT would be a great
advantage:
The difficulty of meeting
deadlines in August, in the presence of business travel
as well as vacations was
noted.
August 1-21: Work on the PEP Draft; Informal meetings as needed.
August 18: Next Meeting with Laura Perini as chair. Review
status of PEP
in the light of the upcoming release of Draft "-1" of the PEP.
Review the action items from the July 30 Meeting, especially
the task/responsibility matrix.
August 21 (?): Distribute Draft "-1", for Comment.
August 27: Meet to finalize comments
on Draft -1 and form Draft "0".
Assign One Editor to collect comments and further develop
the draft.
[NB -- this is during the Atlas SW Week at Ann Arbor.
Virtual room video is available from there, and it
is suggested that we conduct the meeting with
broad Atlas participation from Ann Arbor as well
as CERN.]
Sept. 8
Distribute Draft 0 for Comment
Sept. 14 Meet in
Lyon (Monday morning);
Final comments on Draft 0 discussed;
Form Draft 1 and distribute for comment
Sept. 19 (Saturday)
meeting at CERN for intensive work on the
Draft to be submitted to the LCB
Sept. 22 Submit PEP
to LCB
Sept. 26 Very last
comments on the PEP.
October 1 Final submission of
the PEP to the LCB
October 6 Present PEP
October 7 Project (Officially) Starts.
o Volunteers for writing the PEP were requested.
It was also requested that Krzysztof
begin on the structure of the PEP very soon,
as Laura and Harvey will have
limited availability during the first half of August.
It was agreed that the team of Krzysztof,
Laura, Harvey, and Paolo would coordinate
on the writing. Mauro volunteered for
some writing assignments. CERN/IT contributions
also were solicited, in the areas outlined
above.
o Paolo suggested a solution for MONARC management, satisfying the constraints of
o AOB: The issue of the Modeler (update):
A job description for the Modeler has been finalized:
Modeler/Analyst of
Distributed Computing Systems
for Experiments at
CERN's Large Hadron Collider
Job Description
_________________________________________
Task
____
Development of a system of simulation software that allows the
modeling of globally distributed computer systems for large-scale
physics data analysis.
This work will be performed in the context
of a project being carried out by a collaboration of physicists
and computer scientists from Europe and the US, studying future
Models of how physicists could do the data analysis at experiments
to be performed at CERN's Large Hadron Collider (LHC) starting
in
2005. Data analysis with this scale or degree of geographical
spread
does not exist now, but is expected to become possible through
advances in the information technology over the next several
years.
Qualifications
______________
Education:
________
A university degree in computer science or engineering, physics,
electrical engineering, or the equivalent.
Knowledge and experience:
______________________
The successful candidate will understand how to define software
abstractions of real computers and networks for the purpose
of modeling and simulating their behavior in hypothetical
distributed computing systems.
A proven track record of software development in C++ and/or Java
is required, as is experience and familiarity with software
development on both NT and Unix operating systems. Candidates
should be skilled at quickly understanding existing applications
at the source code level, and be comfortable modifying, developing
and porting those applications.
Candidates must have a good knowledge of modern software design
techniques, GUI-building and especially Object-Oriented analysis
and design. They must understand the need, and be capable of,
fully documenting their work.
Familiarity with an existing (commercial or other) modeling tools,
such as ModSim or ComNet, would be a major advantage.
Good knowledge of English, including the ability to write clear
documentation is required. A working knowledge of French or Italian
is an advantage.
Location:
________
The candidate will be located at CERN, but some travel to the
U.S.
and European countries is foreseen.
Duration:
________
The project will last for a minimum of two years. The candidate's
contract is subject to annual renewal, for the duration of the
project.
Salary:
______
From 5500 CHF to 7000 CHF/month, according to experience and level
of expertise.
-- Harvey Newman
August 1, 1998