GRIDbased data handling for CDFUK

About This Presentation

Title:

Description:

Number of Views:21

Avg rating:3.0/5.0

Slides: 17

Provided by: ianm153

Category:

more less

Transcript and Presenter's Notes

Title: GRIDbased data handling for CDFUK

1
GRID-based data handling for CDF-UK

2
The CDF Experiment

3
Processing/analysis scheme

Reconstruction carried out at Fermilab (FNAL).
Reduces data volume to about 50KB/event. Writes
out Physics Analysis Data (PAD) containing
objects like hits, tracks etc.
PAD data has about 10 streams sorted by Physics
interest.
Each stream grows by typically 6TB/year,
(15GB/day).
UK physicists will be interested in a few of
these streams.
Data analysis runs on PAD data and calculates
physics parameters writing out a brief summary
per event called an NTUPLE (10KB/event)
The final stage of analysis selects and summaries
these NTUPLES (outputs histograms).

4
Processing Requirements in the UK

Code Development and testing
Runs many times
Need to run on databases of approx 30GB
Aim for near interactive performance, 5 mins
100MB/sec
Need to locate datasets containing most
up-to-date or best data
Analysis
Run only a few times
Typically run on large dataset, approx 6TB
Aim for overnight turnaround, 16 hrs 100 MB/sec
Aim to keep at least one stream up-to-date via
network transfers, (15GB/day 1.4Mbps)
Analysis summary
Runs many times
Datasets of a few 100 GB.
Most interactive phase, fastest possible
turnaround required.
Monte Carlo simulated events
Can be very effectively distributed

5
UK Hardware Status

CDF collaborators in the UK applied for JIF grant
for IT equipment in 1998. Awarded 1.67M in
summer 2000.
First half of grant will buy
Multiprocessor systems plus 1TB of disk for 4
Universities
2 multiprocessors plus 2.5 TB of disk for RAL
A 32 CPU farm for RAL (Monte Carlo generation)
5 TB of disk and 8 high end workstations for
FNAL
Emphasis on high IO throughput super-workstations
for end-stage analysis and code
development/testing. Systems will also be used
for running analysis jobs.
A dedicated network link from UK to FNAL

6
CDF-UK JIF Equipment
7
CDF-UK GRID Project

JIF proposal only covered hardware but in the
meantime GRID has arrived !
Use existing tools to make the existing FNAL
system more distributed.
Concentrate on solving real-user issues. Develop
an architecture for locating data, data
replication and job submission .
Make routine tasks automated and intuitive.
Gain experience in GRID and the Globus toolkit.

8
Some Requirements

We must make best use of the existing packages
and minimise modifications to existing software.
Have chosen to emulate the FNAL system so that
analysis code can run without modification. Locks
us into an existing architecture.
Aim to provide a scheme to allow efficient use of
the users time.
Make best use of the systems available by placing
jobs and data for optimal turnaround time.
Produce a simple but useful system ASAP.
Build a system on which more intelligent layers
can be built later.

9
Design principles

All sites are equal
All sites hold meta-data describing only local
data. This stores relationships between files and
datasets, run conditions and luminosity.
Use LDAP to publish meta-data kept in
Oracle - at FNAL
msql also supported by software so use for local
metadata ( mySQL may be possible later)
Use pre and post processing at endpoints of data
transfer
take better account of resources e.g. use of
near-line storage, disk space management
Execute consistency and sanity checks.
Use existing Disk Inventory Manager from FNAL.
Use Globus tools throughout single
authentication.

10
Functionality at a site

Publication of the local metadata
Publication of information about other system
resources (CPU, Disk, Batch queues etc).
Transmission of data via network.
This may involve staging of data from tape to
disk before transmission.
Receive data from the network or from tapes.
Copy or construct metadata
A mechanism to allow jobs from participating
sites to be run.
Some sites may have reduced functionality

11
Data Location/Copy
12
Applications
User Interface
Dataset maintainer
...
Data locator
Data copier
Job Submission
Globus toolkit
13
Scope

Plan to install at
4 UK universities (Glasgow, Liverpool, Oxford,
UCL)
RAL
FNAL (although this would be reduced
functionality, data and metadata exporter)
More non-UK sites could be included
Intend to have basic utilities in place at time
of equipment installation (May 2001)

14
Work so far

Week long meeting with author of existing data
handling software (DIM).
Project plan under development .
UK Universities have agreed to participate.
Expect continuing support from RAL.
Globus installed at a number of sites. Remote
execution of shell commands checked.
LDAP/database connectivity, (S.Fisher, RAL)
LDAP to Oracle via Python script
Java to LDAP (via JNDI) JNDI (Java Naming and
Directory Interface) gives very elegant interface
to LDAP

15
Longer Term plans and needs

Need to understand how to track status of
distributed tasks.
Understand if this strategy is scalable to 100s
of systems. Are there ways to organise LDAP so
that its not too big an overhead ?
Could Monte Carlo generation take advantage of
the same scheme of job placement and data
cataloguing? This is generally even more
distributed (high cpu/IO ratio)
1st phase of equipment purchase will probably be
SMP machines, easy management, single system
image. Must be usable from day 1. With the right
tools, 2nd phase could be more distributed
Need to work on selective expansion of C
objects to minimise CPU overheads.
Expect user Interface to be implemented as Java
application to give platform independence.
Make intelligent UI to automate or suggest
strategies for replicating data/submitting jobs
etc look to generic solutions
Extend UI to provide a more complete working
environment. Integrate histogram presenters etc..
generic solutions may be possible.

16
Effort

Universities finding small fractions of people
adding up to around 0.5 FTE each. Need larger
chunks of more people.
Oxford has created a new technician post which
will free up some effort for GRID.
Italian Computer Science student to join for 6
month project from Easter 2001.
Share experience by involving same staff in both
CDF and other GRID projects ideally full time
GRID staff.

Write a Comment

User Comments (0)

About PowerShow.com

GRIDbased data handling for CDFUK - PowerPoint PPT Presentation