1 BIG FARMS AND THE GRID Job Submission and Monitoring issues ATF Meeting, 20/06/03 Sergio Andreozzi.

Slides:



Advertisements
Similar presentations
Andrew McNab - Manchester HEP - 17 September 2002 Putting Existing Farms on the Testbed Manchester DZero/Atlas and BaBar farms are available via the Testbed.
Advertisements

FP7-INFRA Enabling Grids for E-sciencE EGEE Induction Grid training for users, Institute of Physics Belgrade, Serbia Sep. 19, 2008.
CERN LCG Overview & Scaling challenges David Smith For LCG Deployment Group CERN HEPiX 2003, Vancouver.
Job Submission The European DataGrid Project Team
WP 1 Grid Workload Management Massimo Sgaravatto INFN Padova.
Dr. David Wallom Use of Condor in our Campus Grid and the University September 2004.
A conceptual model of grid resources and services Authors: Sergio Andreozzi Massimo Sgaravatto Cristina Vistoli Presenter: Sergio Andreozzi INFN-CNAF Bologna.
GRID Workload Management System Massimo Sgaravatto INFN Padova.
Grid Computing, B. Wilkinson, 20046c.1 Globus III - Information Services.
EDG Resource Broker for the Glue Schema Sergio Andreozzi INFN-CNAF Bologna (Italy)
Pilots 2.0: DIRAC pilots for all the skies Federico Stagni, A.McNab, C.Luzzi, A.Tsaregorodtsev On behalf of the DIRAC consortium and the LHCb collaboration.
GLUE Schema: conceptual model and implementation
Workload Management WP Status and next steps Massimo Sgaravatto INFN Padova.
EGEE-II INFSO-RI Enabling Grids for E-sciencE EGEE Middleware: gLite Information Systems (IS) EGEE Tutorial 23 rd APAN Meeting,
A. Cavalli - F. Semeria INFN Experience With Globus GIS 1 A. Cavalli - F. Semeria INFN First INFN Grid Workshop Catania, 9-11 April 2001 INFN Experience.
INFSO-RI Enabling Grids for E-sciencE Logging and Bookkeeping and Job Provenance Services Ludek Matyska (CESNET) on behalf of the.
Grid Workload Management & Condor Massimo Sgaravatto INFN Padova.
CHEP 2003Stefan Stonjek1 Physics with SAM-Grid Stefan Stonjek University of Oxford CHEP th March 2003 San Diego.
1 st December 2003 JIM for CDF 1 JIM and SAMGrid for CDF Mòrag Burgon-Lyon University of Glasgow.
DataGrid WP1 Massimo Sgaravatto INFN Padova. WP1 (Grid Workload Management) Objective of the first DataGrid workpackage is (according to the project "Technical.
INFSO-RI Enabling Grids for E-sciencE Workload Management System Mike Mineter
A monitoring tool for a GRID operation center Sergio Andreozzi (INFN CNAF), Sergio Fantinel (INFN Padova), David Rebatto (INFN Milano), Gennaro Tortone.
Grid Workload Management Massimo Sgaravatto INFN Padova.
LCG / ARC Interoperability Status Michael Grønager, PhD (UNI-C / NBI) January 19, 2006, Uppsala.
Grid job submission using HTCondor Andrew Lahiff.
Andrew McNabETF Firewall Meeting, NeSC, 5 Nov 2002Slide 1 Firewall issues for Globus 2 and EDG Andrew McNab High Energy Physics University of Manchester.
Mar 28, 20071/18 The OSG Resource Selection Service (ReSS) Gabriele Garzoglio OSG Resource Selection Service (ReSS) Don Petravick for Gabriele Garzoglio.
Open Science Grid OSG CE Quick Install Guide Siddhartha E.S University of Florida.
EGEE-III INFSO-RI Enabling Grids for E-sciencE Feb. 06, Introduction to High Performance and Grid Computing Faculty of Sciences,
CERN Using the SAM framework for the CMS specific tests Andrea Sciabà System Analysis WG Meeting 15 November, 2007.
EGEE is a project funded by the European Union under contract INFSO-RI Copyright (c) Members of the EGEE Collaboration GLUE Schema Sergio.
What is SAM-Grid? Job Handling Data Handling Monitoring and Information.
Getting started DIRAC Project. Outline  DIRAC information system  Documentation sources  DIRAC users and groups  Registration with DIRAC  Getting.
INFSO-RI Enabling Grids for E-sciencE OSG-LCG Interoperability Activity Author: Laurence Field (CERN)
Enabling Grids for E-sciencE SGE J. Lopez, A. Simon, E. Freire, G. Borges, K. M. Sephton All Hands Meeting Dublin, Ireland 12 Dec 2007 Batch system support.
Proposal for a IS schema Massimo Sgaravatto INFN Padova.
E-infrastructure shared between Europe and Latin America FP6−2004−Infrastructures−6-SSA gLite Information System Pedro Rausch IF.
1 Andrea Sciabà CERN Critical Services and Monitoring - CMS Andrea Sciabà WLCG Service Reliability Workshop 26 – 30 November, 2007.
1 Grid2Win: porting of gLite middleware to Windows Dario Russo INFN Catania
INFSO-RI Enabling Grids for E-sciencE GridICE: Grid and Fabric Monitoring Integrated for gLite-based Sites Sergio Fantinel INFN.
LCG workshop on Operational Issues CERN November, EGEE CIC activities (SA1) Accounting: current status
High-Performance Computing Lab Overview: Job Submission in EDG & Globus November 2002 Wei Xing.
DataTAG is a project funded by the European Union DataTAG WP4 meeting, Bologna 29/07/2003 – n o 1 GLUE Schema - Status Report DataTAG WP4 meeting Bologna,
GIIS Implementation and Requirements F. Semeria INFN European Datagrid Conference Amsterdam, 7 March 2001.
EDG - WP1 (Grid Work Scheduling) Status and plans Massimo Sgaravatto INFN Padova.
Open Science Grid Build a Grid Session Siddhartha E.S University of Florida.
Grid Workload Management (WP 1) Massimo Sgaravatto INFN Padova.
EGEE is a project funded by the European Union under contract INFSO-RI Grid accounting with GridICE Sergio Fantinel, INFN LNL/PD LCG Workshop November.
EGEE 3 rd conference - Athens – 20/04/2005 CREAM JDL vs JSDL Massimo Sgaravatto INFN - Padova.
Gennaro Tortone, Sergio Fantinel – Bologna, LCG-EDT Monitoring Service DataTAG WP4 Monitoring Group DataTAG WP4 meeting Bologna –
DataTAG is a project funded by the European Union International School on Grid Computing, 23 Jul 2003 – n o 1 GridICE The eyes of the grid PART I. Introduction.
DataTAG is a project funded by the European Union CERN, 8 May 2003 – n o 1 / 10 Grid Monitoring A conceptual introduction to GridICE Sergio Andreozzi
Consorzio COMETA - Progetto PI2S2 UNIONE EUROPEA Grid2Win : gLite for Microsoft Windows Elisa Ingrà - INFN.
INFN/IGI contributions Federated Clouds Task Force F2F meeting November 24, 2011, Amsterdam.
Integrating HTCondor with ARC Andrew Lahiff, STFC Rutherford Appleton Laboratory HTCondor/ARC CE Workshop, Barcelona.
CE design report Luigi Zangrando
DIRAC: Workload Management System Garonne Vincent, Tsaregorodtsev Andrei, Centre de Physique des Particules de Marseille Stockes-rees Ian, University of.
Information System testing for LCG-1
Dag Toppe Larsen UiB/CERN CERN,
Dag Toppe Larsen UiB/CERN CERN,
Practical: The Information Systems
Towards GLUE Schema 2.0 Sergio Andreozzi INFN-CNAF Bologna, Italy
Resource monitoring and discovery in OGSA
Sergio Fantinel, INFN LNL/PD
Developments in Batch and the Grid
CMS report from FNAL demo week Marco Verlato (INFN-Padova)
gLite Information System
A conceptual model of grid resources and services
EGEE Middleware: gLite Information Systems (IS)
Information System (BDII)
Presentation transcript:

1 BIG FARMS AND THE GRID Job Submission and Monitoring issues ATF Meeting, 20/06/03 Sergio Andreozzi

2 OUTLINE Computing Resources in the Glue Schema  CE is access point to a queue (same as previous EDG)  only one access point envisioned per queue  E.g. edt001.cnaf.infn.it:2119/jobmanager-pbs-short Typical Cluster Configuration Broker Service and Cluster Services  Description of a the interaction  (goal: understand the submission process and how cluster services are stressed) LCG Proposed Cluster Configuration Job submission in this new scenario Discussion Proposal

3 Typical cluster layout In DataGrid, the typical cluster configuration is:  One node running: Batch server (e.g. LSF, PBS) MDS GRIS Gatekeeper  Several Worker nodes driven by the batch server for job execution

4 Typical cluster layout Worker node Worker node Worker node Worker node … Head node Batch server gatekeepergris queue

5 Typical job submission (partial description) Head Node edt001 Batch server gatekeepergris Broker Information Index Head Node edt002 Batch server gatekeepergris 2 Two different clusters 1.From II, broker gets list of CEs a user can access and that match JDL requirements 2.For each selected CE, the GRIS is contacted to get params in JDL rank option to order 3.The first CE in ordered list is used to run the job (I don’t remember if requirements are checked in 1 or 2)

6 Cluster CERN assert:  Gatekeeper service can be heavily loaded when managing several job submissions e.g. gatekeeper design issue; one living process for each submitted job till the end of the computation  For scalability they deploy gatekeeper service in a different node than batch server  They can have several nodes running gatekeeper for the same batch server  They plan to set up a big farm with nodes with a set of O(10) access nodes and only one batch server  LSF can manage O(1000) nodes

7 Cluster CERN Worker node Worker node Worker node Worker node … Head node Batch server Access node gatekeepergris Access node gatekeepergris Access node gatekeepergris … queue

8 Example Head node Batch server LSF Access node edt001 gatekeepergris Access node edt002 gatekeepergris Access node edt00n gatekeepergris … Broker shlo CE:edt001:2119/sh CE:edt001:2119/lo CE:edt002:2119/sh CE:edt002:2119/lo CE:edt00n:2119/sh CE:edt00n:2119/lo Each CE is a different queue for the broker Load balancing made among matching queues Among replicated queues, the rank process always provide the same order Information index

9 Cluster CERN Adv & Disadv +  Can scale to higher number of parallel job submission  From the site manager viewpoint, it provides great flexibility in managing/configuring the farm -  Not envisioned in both EDG schema and Glue schema -> duplication of info E.g. given a queue/CE on a batch server, this will show in the GIS as many times as the number of configured gatekeepers for the batch server managing the queue  The LRMS is stresses from several info providers asking same info; e.g. 10 gatekeepers/GRIS for an LSF server refreshing info each 30s 20 req each min instead of 2… this might be a problem

10 Discussion How does the broker deal with this scenario? What are the needed changes to support this?  GIS schema introducing the concept of Access Point; a queue can have several Access Point Defining a quality parameter for an access point (so that the broker can choose the less loaded one)  broker service  monitoring service

11 GRIS AND GATEKEEPER SERVICES The important question is:  DO WE NEED TO REPLICATE THE GRIS AS WELL? The gakepeer does not need the GRIS The GRIS at the moment need some info from the gatekeeper  e.g. hostname, port, access to gridmap file If we don’t miss anything else, they can be easily decoupled on different machines

12 PROPOSAL Decouple the CE Unique ID from the entry point CE ID should be just a global unique ID for the queue E.g.  /jobmanager- - Introduce a new attribute GlueCEAccessPoint  E.g. = current GlueCEUniqueID ONE GRIS PER BATCH SERVER

13 Proposed scenario Worker node Worker node Worker node Worker node … Head node Batch server Access node gatekeeper Access node gatekeeper Access node gatekeeper … queue gris Can run on an access node, on the Head node or on another machine ONLY ONE ISTANCE

14 EXAMPLE CE REPRESENTATION WITHIN MODIFIED GLUE SCHEMA FOR LDAP dn: GlueCEUniqueID=edt001.cnaf.infn.it/jobmanager-pbs-short, Mds-Vo-Name=local, o=grid... GlueCEUniqueID: edt001.cnaf.infn.it/jobmanager-pbs-short GlueCEAccessPoint: edt002.cnaf.infn.it:2119/jobmanager-pbs-short GlueCEAccessPoint: edt003.cnaf.infn.it:2119/jobmanager-pbs-short GlueCEAccessPoint: edt004.cnaf.infn.it:2119/jobmanager-pbs-short GlueCEStateFreeCPUs: 5 GlueCEPolicyMaxRunningJobs: 10 GlueCEAccessControlBaseRule: … …

15 Broker modification When querying the GRIS, the broker will maintain the several Access Points Once it selects the queue, it will submit to a random gatekepeeper Further improvement*: GlueCEAccessPoint: edt004.cnaf.infn.it:2119/jobmanager-pbs- short: Where CE_AP_LOAD is a metric that let the broker able to rank the access points BENEFITS:  the broker will do less work during matchmaking process (no duplicated GRIS to be queried to get info for ranking on EACH JOB SUBMISSION)  the broker will choose a random gatekeer among the listed as it does now (so no worse behaviour); this can be improved*

16 Monitoring No need to deal with replicated info Able to show real number of queues, their state and access points Detailed host info can be aggregated and presented as access node loads