Документ взят из кэша поисковой машины. Адрес
оригинального документа
: http://www.adass.org/adass/proceedings/adass00/P2-39/
Дата изменения: Tue May 29 20:05:47 2001
Дата индексирования: Tue Oct 2 05:52:32 2012
Кодировка:
Поисковые слова: р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п р п п п п п п п п п п п п п п п п
|
Next: A High Throughput Photometric Pipeline
Up: Science Data Pipelines
Previous: Automated Reduction and Calibration of SCUBA Archive Data Using ORAC-DR
Table of Contents -
Subject Index -
Author Index -
Search -
PS reprint -
PDF reprint
Subramanian, S. & Plummer, D. A. 2001, in ASP Conf. Ser., Vol. 238, Astronomical Data Analysis Software and Systems X, eds. F. R. Harnden, Jr., F. A. Primini, & H. E. Payne (San Francisco: ASP), 303
The Chandra Automatic Processing/Archive Interface
Sreelatha Subramanian, David Plummer
Harvard-Smithsonian Center for Astrophysics,
60 Garden St. MS-81, Cambridge, MA 02138,
Email: latha@head-cfa.harvard.edu
Abstract:
The Chandra Automatic Data Processing System (AP) requires quick access
to previously generated data. Potential inefficiencies are avoided by
introducing a layer between the pipelines and the archive. This archive
interface layer includes an archive request queue, a data archiving
server (darch), and an archive ``cache''. The design and functional
operation of each of these components are presented in this paper.
The first step in improving the AP system is the implementation of a cache.
In theory, the cache operates as a typical cache--data is retrieved from a
slow-access storage facility and is stored to a higher-speed retrieval
location, with the expectation that it will be used again.
In our implementation, the cache is actually a directory where the data files
reside, as shown in Figure 1. While a pipeline is processing, it will
store its data products to the cache directory. These data products can then
be stored to the archive at any time, separating data product ingestion from
processing. When another pipeline requires those data products, the cache is
checked first. If the data products are not in the cache, then the pipeline
retrieves the files from the archive database. This method is especially
significant when processing data for the first time since the system can
access all the data products from cache, avoiding the archive database. If
re-processing is required, data will be retrieved from the archive if it is
no longer in cache.
Figure 1:
AP/Archive Interface
|
The implementation of a cache introduces the issue of memory usage. The cache
server loads all the files in the cache directory into its memory, which
allows for quick search and retrieval of the data. However, as the size
of the cache increases, the cache server requires more memory. To reduce the
number of files in cache, an inactive cache directory structure is created.
The cache server does not load files that exist in the inactive directory into
its memory. It stores a few pieces of information about the file along with
the filename, in order to facilitate retrieval. The implementation of the
inactive cache eliminates the issue of memory usage, while maintaining the
usefulness of the cache.
A significant improvement to the cache system is made through the
implementation of virtual files. In our system, virtual files are
simply files that do not exist in the cache, but whose filename and
metafile still exist. With this information, the cache can be queried to
retrieve a file based on time, data product id or other information.
The cache returns the filenames of the files it found, and also returns an
error code which indicates that the files are virtual. The AP system,
recognizing that the retrieved files are virtual, can then use the
filename to retrieve the file from the archive database. This
modification is significant because retrieving a file from the archive
using the filename as the key is more efficient than retrieving using
another key (such as data product id, or time).
The other important
factor is that the implementation of virtual files allows us to search
for file(s) by using cache specific tag information. Tag information,
which is stored in the metafile, is not entered into the archive.
Therefore it was previously not possible to retrieve files from the
archive using tag information. With virtual files, however, this
becomes possible. A search by tag function searches the cache, using the
tag. As this information is in the metafile, the function will return
a corresponding filename, which can then be used to retrieve the file
from the archive.
Another important modification to the AP System is the implementation of file queues.
File queues are implemented to eliminate the time a pipeline must wait for its
data products to be ingested into the archive database. Instead of waiting for
a file to be ingested, the ingestion request is sent to a file queue. The request
is then picked up at a later time by the data archiving server, or darch, and
processed. The implementation of file queues is simply a file which contains all
the requests. The pipeline will not need to wait for ingestion since it has
already stored the file to cache. Therefore, ingestion of the file does not
need to occur simultaneously, and the file queue allows for another process to
handle the requests at any time.
The archive interface layer was not part of the original design. However, the
necessity for it was identified during the processing of ground calibration data
of the system before launch. This allowed us to implement and integrate the
archive interface layer into the AP system, to avoid possible bottlenecks during
the mission.
This project is supported by the Chandra X-ray Center under NASA contract
NAS8-39073.
© Copyright 2001 Astronomical Society of the Pacific, 390 Ashton Avenue, San Francisco, California 94112, USA
Next: A High Throughput Photometric Pipeline
Up: Science Data Pipelines
Previous: Automated Reduction and Calibration of SCUBA Archive Data Using ORAC-DR
Table of Contents -
Subject Index -
Author Index -
Search -
PS reprint -
PDF reprint
adass-editors@head-cfa.harvard.edu