HI Thomas, the current version of rootd serves just byte ranges from a file without any additional intelligence concerning the contents of the file. It interfaces on the lowest level of TFile (seek, readbuffer, writebuffer, etc.). It is possible to implement higher level protocols where complete events are transfered via TMessages. But this is a bit more tricky if it involves user objects that must be known on both the client and remote site. Tricky in the sense as to make a generic solution. However, implementing such a "custom" server is fairly easy for specific cases. The best solution to your problem is PROOF and which we are currently working. Cheers, Fons. USENET News System wrote: > > >From news Fri Aug 4 17:45 MET 2000 > Path: cern.ch!claire.desy.de!not-for-mail > From: Thomas Hadig <hadig@rwth-aachen.de> > Newsgroups: cern.root,desy.root > Subject: rootd and event-wise information > Date: Fri, 04 Aug 2000 17:44:44 +0200 > Organization: DESY > Lines: 41 > Message-ID: <398AE4EC.51BBFBC5@rwth-aachen.de> > NNTP-Posting-Host: tanagra.desy.de > Mime-Version: 1.0 > Content-Transfer-Encoding: 7bit > X-Mailer: Mozilla 4.7 [en] (X11; I; Linux 2.2.12 i686) > X-Accept-Language: en > Xref: cern.ch cern.root:7970 desy.root:36 > Content-Type: text/plain; charset=us-ascii > Content-Length: 2034 > > Dear all, > > i was wondering about the performance of network access to ROOT files. > > In a normal analysis job I can see two major access patterns: > * Sequential access to all events. This will happen in case of already > selected events, special Monte Carlo productions and during scanning > tag files. > * Access to a small percentage of events in a file. This will occur > on large data sets that are preselected by e.g. a tagging mechanism. > > In the latter part, the data will be accessed in blocks (e.g. a TBasket > in ROOT files). In case you access only a single event from a basket, > this will mean that a large overhead is sent over the network and will > reduce the performance for non-local data files. > > In case you always read the full information of an event, it might be > possible to make an event-wise information stream that sends one event > over the network and not one block. In case of splitted events where > you read only the variables you need, this will not be possible since > it is a priori not known what part of the information should be send. > In that case also the latency of the network might play a considerable > role in processing speed. > > I know that an improved network protocol is implemented in rootd coming > with ROOT. However, i could not find any information on the Web pages > on how rootd tackles this problem and what real-life experiences with > such data storage setups are. > > Ciao and thanks in advance > Thomas > > ---------------------------------------------------------------------------- > Thomas Hadig, Ottensen, Hospitalstr. 126, 22767 HH Tel. +49 40 39901186 > OFFICE : DESY/H1, 1a334, Notkestr. 85, 22607 Hamburg Tel. +49 40 8998 2733 > hadig@rwth-aachen.de Handy +49 170 7807434 > http://www-h1.desy.de/~hadig/ FAX +49 40 8998 4385 > ---------------------------------------------------------------------------- > In a world without walls and fences, who needs windows and gates ? > ---------------------------------------------------------------------------- -- Org: CERN, European Laboratory for Particle Physics. Mail: 1211 Geneve 23, Switzerland E-Mail: Fons.Rademakers@cern.ch Phone: +41 22 7679248 WWW: http://root.cern.ch/~rdm/ Fax: +41 22 7677910
This archive was generated by hypermail 2b29 : Tue Jan 02 2001 - 11:50:31 MET