Post on 21-Dec-2015
transcript
Grid-Brick Event Processing Framework in GEPS
CHEP 03 – La Jolla, California
A. Amorim, P. Trezentos, N. Almeida, H. Fei, L.Pedro, J.Villate, H.Wolters
2
Outline Introduction Architecture Overview Data Flow How do we do it Action! Vantages and disadvantages On-going and future work Summary
luis.pedro@cern.ch FCUL - Lisbon
3
Introduction
What is GEPS Grid-based Event Processing System Developed on top of Globus Provides web-based access to grid
computing environment for event processing Developed by GridPT working group
luis.pedro@cern.ch FCUL - Lisbon
4
Introduction
Main Idea
Do NOT move data!
Each node stores and is
reponsible by a subset of the whole data…
luis.pedro@cern.ch FCUL - Lisbon
6
Introduction
Our way
User terminal User terminal User terminal………..
Job submit server
Meta-data catalogue
luis.pedro@cern.ch FCUL - Lisbon
7
Architecture Overview
User submits a query through a interface to the Job Submit Server (JSS).
Job submitted information will be stored in the Metadata Catalog.
luis.pedro@cern.ch FCUL - Lisbon
8
Architecture Overview
luis.pedro@cern.ch FCUL - Lisbon
The job is submitted to the grid nodes using Globus API functions.
All the nodes query their own information and retrieve a result.
9
Architecture Overview
JSS receives result from the Grid nodes and produces a final one
User can download or consult the final result
User can also visualize the state of the job in each Grid node
luis.pedro@cern.ch FCUL - Lisbon
11
Data-flow
User terminal(PHP interface)
BROKER
Meta-data catalogue
luis.pedro@cern.ch FCUL - Lisbon
JSS
Grid node
Brick
13
How do we do it:
Relevant Features Globus
Toolkit that provide GRID API functions
PgSQL Meta-data catalogue implementation
LDAP Query Grid node information
PHP Web interface
luis.pedro@cern.ch FCUL - Lisbon
14
Action!
The human interface
Main Page
Submit a Job
GREED info
Job status
luis.pedro@cern.ch FCUL - Lisbon
15
Action!
Enabling ROOT Queries The job is submitted to the grid nodes All the nodes query their own information with
ROOT and retrieve a ROOT file with a TTree JSS receives the ROOT files and produces a
final ROOT file with the result of the query User can download or consult the final file
because it is a TTree.
luis.pedro@cern.ch FCUL - Lisbon
16
Action!
Enabling ROOT Queries (cont) Stores the information in each node using TTree’s Filter the information in each node and retrieve a
result file that include a TTree Join all the result files in the Job Submit Server using
a TChain and produces a final TTree that is the query result
View the final result file with a TBrowser or with Carrot
luis.pedro@cern.ch FCUL - Lisbon
17
Action!
Enabling ROOT Queries (reading ROOT files)
Analysis Performance Reading From a TFile
050
100150200250300350400
100 500 1000 2000 4000
Events
Tim
e S
pen
t (s
)
30% Events
60% Events
90% Events
luis.pedro@cern.ch FCUL - Lisbon
18
Action!
Enabling ROOT Queries (reading ROOT files)
Analysis Performance Reading From a TTree
0
20
40
60
80
100
120
140
100 500 1000 2000 4000 6000
Events
Tim
e S
pen
t (s
)
30% Events
60% Events
90% Events
luis.pedro@cern.ch FCUL - Lisbon
19
Action!
Special Features used ROOT
TObject TTree CINT Filtering data from TTree ROOT I/O TChain
Carrot Browsing ROOT files Histograming variables
luis.pedro@cern.ch FCUL - Lisbon
20
Vantages and disavantages
Vantages Commodity Data Storage Huge Scalability (400 GB/node) Granularity
Disadvantages Load balancing
Suitable storage policy Fault tolerance
Data replication or Backup
luis.pedro@cern.ch FCUL - Lisbon
21
On-going and future work
Error handling and fault-tolerance Recover mechanisms for each node Create a redundancy mechanism to recover
from a malfunction in the nodes Develop a storage mechanism to submit more
work to the best nodes Load balancing
Provide to user several interfaces to submit work
luis.pedro@cern.ch FCUL - Lisbon
22
Summary
A different approach is being developed There is already a real prototype working Some (good) results have been achieved A lot of work already done…..
But …. Still a lot of work to do!
luis.pedro@cern.ch FCUL - Lisbon
23
Acknowledgments
Thank you to those who are developing and participating in this project A. Amorim (Antonio.Amorim@fc.ul.pt) P. Trezentos (Paulo.Trezentos@iscte.pt) N. Almeida (Nuno.Almeida@iscte.pt) H. Fei (hnfi@iscte.pt) L.Pedro (Luis.Pedro@cern.ch) J.Villate (villate@fe.up.pt) H.Wolters (helmut@lip.pt)
Keep the good workThanks for hearing me!
luis.pedro@cern.ch FCUL - Lisbon