infso-ri-508833 enabling grids for e-science planck simulations status of the application c....

28
INFSO-RI-508833 Enabling Grids for E-sciencE www.eu-egee.org Planck Simulations Status of the Application C. Vuerli, G. Taffoni, A. Barisani, A. Zacchei, F. Pasian INAF – Information Systems Unit and OA Trieste Geneve, 1 March 2006

Upload: griselda-armstrong

Post on 30-Dec-2015

215 views

Category:

Documents


2 download

TRANSCRIPT

Diapositiva 1C. Vuerli, G. Taffoni, A. Barisani, A. Zacchei, F. Pasian
INAF – Information Systems Unit and OA Trieste
Geneve, 1 March 2006
Wednesday 1 March 2006
The Grid added value
Future perspectives and issues for use of Grid Technology
Summing up status
Timeline
(a TOD of ~7 TB for the whole LFI mission)
changing calibration (parameters configuration)
Wednesday 1 March 2006
Brief introduction
Goal: make possible N simulations of the whole Planck/LFI mission (~14 months), each time with different cosmological and instrumental parameters
Full sky maps production for frequencies 30-857 GHz by means of two complete sky surveys
Sensitivity of a few μK per pixel 0,3° in amplitude
22 channels for LFI, 48 for HFI
Data volume produced at the end of the mission: ~2TB for LFI and ~15TB for HFI
Computing requirements: ~100Tflops for raw data reduction, foregrounds extraction and CMB maps creation
Wednesday 1 March 2006
(43 executables and a few libraries)
Languages used are C/C++/Fortran/F90;
Shell/Perl for scripts;
Porting of Monte Carlo simulation code by Sam Leach
Planck simulation is a set of 70 instances of the Pipeline (22 for LFI and 48 for HFI)
Wednesday 1 March 2006
Is it parallel? NO it runs concurrently.
Do we need MPI/parallel? Yes. In later phase for data analysis (16/32 CPUs in the site).
Do we produce data? YES, we have an intensive data production. Can be more than 1 PB.
How long does it run? From 6h (short) up to 36h (long)
Access to/exchange of data coming from other experiments (IVOA, MAGIC)
Wednesday 1 March 2006
Replica and security;
data.
Planck simulations are highly computing demanding and produce a huge amount of data. Such resources cannot be usually afforded by a single research institute, both in terms of computing power and data storage space.
Wednesday 1 March 2006
Flexible environment;
Collaborative work for simulations and reduction:
less time, less space, less frustration….
VObs view:
Native authentication/authorization mechanism;
A federation of users within a VO fosters the scientific collaboration;
Collaborative work between different applications.
Wednesday 1 March 2006
First Tests
First tests performed on a workstation aimed at identifying the computational and storage needs of the simulation SW in detail
Computational time on a dual CPU 2.4 GHz workstation with 2 GB of RAM
for the whole simulation of the LFI mission
[4 radiometers at 30 GHz, 6 at 44 GHz and 12 a 70 GHz]
LFI 30GHz
LFI 44GHz
LFI 70GHz
Grid Environment
To allow users to run Planck simulation SW we need to create some specific services on Grid general environment;
They must be used to run both one or more SDPs (Single Detector Pipeline) or the whole MSJ (Mission Simulation Job);
They are modular and easy to integrate with new pipeline stages when some upgrade is needed (this is necessary if we want to allow users to develop new codes).
Wednesday 1 March 2006
Grid Environment
Step 1: Deployment of the simulations code on the Grid as RPMs. In this first test however we used the Replica Manager to copy and register the SW.
Step 2: Creation of an application specific environment on top of the UI: a set of Perl scripts are available allowing a user to configure a pipeline and submit it to the Grid.
Step 3: Implementation of a metadata description to identify the cosmological and instrumental parameters and to associate them to the GUID of a complex output file (TODs, maps, noise contributions etc.)
Important for post processing analysis.
Wednesday 1 March 2006
Simulations description
A number of test simulations of the MSJ run with the same parameters used on the dual-CPU WS;
Initially selected only the Grid sites equipped with Xeon WNs at CPU speed of ~2400 and sites with at least 1 free CPU;
Run sets of MSJ with different degree of parallelization;
Tests repeated 30 times under different load conditions of the Grid to verify the stability of both the submission tools and of the Grid environment.
We noticed that RB usually assigns each SDP to a different site, so the MSJ runs on a truly distributed environment. However, a few times the jobs were assigned to the same site but to different WNs. As expected, no significant benefit or decay in performance was noticed in those cases. Also different Grid load did not change significantly the results.
Wednesday 1 March 2006
Simulations description
Set of tests involving the whole computing power and data storage available to our VO (~5000 CPUs of different kind). Submission of 100 concurrent MSJs from different UIs with the only requirement of finding enough free disk space to save the output.
Tests span two years starting from summer 2004 and using different versions of the MW (2.2, 2.4 and 2.6) and within different VOs
The whole test lasts for ~3 days and was repeated different times under different load conditions of the Grid with no significant change in the results
Long simulations could require to modify CFTSIO to allow I/O directly on Grid through GridFTP
Wednesday 1 March 2006
Post-processing example
To verify the possibility of using the stored TODs for some kind of post-processing we applied the destriping algorithm on the TODs produced during the short runs.
Metadata are used to locate the files and to identify their GUIDs.
The configuration/submission Planck tools are modified to create a JDL with the input-file option that points to the TOD GUID.
The input-file option is used by the RB to force the job to run in the Grid site where the input data are stored. This optimizes the data transfer which is in this way restricted to the site LAN.
The Grid "configurator" is modified to download any input data file specified in the input-file option before running the pipeline.
Wednesday 1 March 2006
Post-processing example
The “destriping” procedure runs for ~20 minutes for 30 GHz channel up to ~40 minutes for the 70 GHz channel on a dual AMD workstation.
On Grid the run time for a simulations set of 22 radiometer is ~55 minutes with a gain of a factor of 10 in performances compared with times required on the workstation.
Wednesday 1 March 2006
Workstation
Grid
Gain
short
ProC and G-DSE
ProC is a scientific workflow engine developed in the framework of IDIS (Integrated Data and Information System) collaboration
It executes “pipelines” of modules
Workflows, directed “acyclic” graphs
It allows the assembly of Pipelines using building block modules
Modules may be heterogeneous (FORTRAN, C, C++, Java, Python, GDL/IDL, ...); also sub-pipelines
It is a data-driven, forward-chaining system
It has components for ...
checking for consistency & completeness
Execution
The G-DSE makes of databases a new embedded resource of the Grid. It enables a new Grid QE to access databases and use them for data analysis (see presentation by G. Taffoni on Thu March 2nd, 2006, at 2.00 PM “Data Access on the Grid” session)
Wednesday 1 March 2006
UI in each site;
Quantum-Grid in each site;
Regional Area
Current status
Future status
R.A.: Italy
15 CPUS 300 GB
15 + 45 CPUS 1 TB (total). All INAF sites are now in the startup phase
R.A.: Spain
R.A.: UK
R.A.: Germany
R.A.: The Netherlands
Wednesday 1 March 2006
Basic gridification of the Application OK
Main issues met in 2005
Slow startup process of Planck VO
Slow start up of interactions between Planck VO site managers and national ROCs
Some technical initial problems (e.g. VOMS)
The management of the VO has proved to be more complex with respect our expectations
Heterogeneous VO
Debugging
The gridification of Planck pipelines has still to be completed
Until now extensive tests involving all nodes of the VO were not possible
Corrective action
On-site (at VO sites) meetings and training events (involving Planck and INAF VOs) addressed to site managers and users and scheduled for the next months for a:
Fast startup of VO nodes with new shared resources available;
Gridification of new pipelines;
Extensive tests within the VO.
Future strictly dependent from a number of factors: gLite (!?!), support (?), EGEE-2 (?)
Wednesday 1 March 2006
Wednesday 1 March 2006