From k.bos@nikhef.nl Mon Feb 5 16:42:07 2001 Date: Mon, 5 Feb 2001 15:14:16 +0100 From: Kors Bos To: a03@nikhef.nl, david.groep@nikhef.nl Cc: bosk@nikhef.nl Subject: use case follow up [The following text is in the "iso-8859-1" character set] [Your display is set for the "US-ASCII" character set] [Some characters may be displayed incorrectly] The example in Germans paper is very much inspired by the CERN setup. Our example is very much inspired by our D0 setup. We would like to go towards a more integrated setup: one farm & one farm server? We continue the use case. The Work Load Manager (WP1) has a request to process 10,000 events. It learns from the Information Server (WP3) that there are 2 centra available for this work. One of those is NIKHEF. We have 100 cpu's. The other centre also has 100 cpu's. The WLM then splits the job in 100 jobs of 50 events to be sent to us and another set like that to be split to the other centre. The Farm Control Manager FCM is the protal to the grid world which receives this request in the Job Definition Language. The FCM will immedeately expand this request into 100 jobs of 50 events. It will then go and fetch the data at the location specified in the JDL. This gridftp filename is then translated into a local filename which can be accessed by the nodes. In a second iteration we discuss this translation in a bit more detail. Suppose the file is specified within the program in a C open statement like open( .. 'gsiftp://grid.cern.ch/lhcb/erikvanherwijnen/inputfile1' ...) When the node where the program is running sees this line it will try to open a port to get this file. We could imagine installing a proxy service on the farm server which catches this request and re-directs it to a file which is already present in the cache. The file has to be declared as well in the JDL as a file needed for the program; otherwise it cannot b cached beforehand. This is WP2's job supposingly. For the outputfile(s) the inverse proces takes place. We could think about using IPV6 on the nodes to make them accessible for monitoring and interactive debugging. The specifications of the job in the JDL has to be translated to the language which can be recognised by the local batch system on the farm server. For the 9-month release we should probably accept just to have a file with names of users. This user name is translated to a local username which only exists for the duration of the job on the fabric. Kors Bos, NIKHEF, Amsterdam, The Netherlands P.O.Box 41882, NL-1009 DB Amsterdam Tel. +31-(0)20-5925083 Gsm. +31-(0)6-53649224 http://www.nikhef.nl/~bosk