All Downloads are FREE. Search and download functionalities are using the official Maven repository.

com.bigdata.bop.engine.IChunkAccessor Maven / Gradle / Ivy

/**

Copyright (C) SYSTAP, LLC DBA Blazegraph 2006-2016.  All rights reserved.

Contact:
     SYSTAP, LLC DBA Blazegraph
     2501 Calvert ST NW #106
     Washington, DC 20008
     [email protected]

This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
*/
/*
 * Created on Sep 13, 2010
 */

package com.bigdata.bop.engine;

import java.nio.ByteBuffer;
import java.util.concurrent.BlockingQueue;

import com.bigdata.bop.IBindingSet;
import com.bigdata.relation.accesspath.BlockingBuffer;
import com.bigdata.striterator.IChunkedIterator;

import cutthecrap.utils.striterators.ICloseableIterator;

/**
 * API providing a variety of ways to access chunks of data (data are typically
 * elements or binding sets).
 * 
 * TODO This could be much more flexible and should be harmonized to support
 * high volume operators, GPU operators, etc. probably the right thing to do is
 * introduce another interface here with a getChunk():IChunk where IChunk let's
 * you access the chunks data in different ways (and chunks can be both
 * {@link IBindingSet} []s and element[]s so we might need to raise that into
 * the interfaces and/or generics as well).
 * 
 * TODO It is likely that we can convert to the use of {@link BlockingQueue}
 * instead of {@link BlockingBuffer} in the operators and then handle the logic
 * for combining chunks inside of the {@link QueryEngine}. E.g., by scanning
 * this list for chunks for the same bopId and combining them logically into a
 * single chunk.
 * 

* For scale-out, chunk combination will naturally occur when the node on which * the operator will run requests the {@link ByteBuffer}s from the source nodes. * Those will get wrapped up logically into a source for processing. For * selective operators, those chunks can be combined before we execute the * operator. For unselective operators, we are going to run over all the data * anyway. * * TODO Expose an {@link IChunkedIterator}, which handles both element at a time * and chunk at a time. * * TODO Expose a mechanism to visit the direct {@link ByteBuffer} slices in * which the data are stored. For an operator which executes on a GPU, we want * to transfer the data from the direct {@link ByteBuffer} in which it was * received into a direct {@link ByteBuffer} which is a slice onto its VRAM. * (And obviously we need to do the reverse with the outputs of a GPU operator). * * @see Optimize * serialization for query messages on cluster * * @author Bryan Thompson * @version $Id$ */ public interface IChunkAccessor { /** * Visit the binding sets in the chunk. */ ICloseableIterator iterator(); // /** // * Chunked iterator pattern. The iterator may be used for element at a time // * processing, but the underlying iterator operators in chunks. The size of // * the chunks depends originally on the data producer, but smaller chunks // * may be automatically combined into larger chunks both during production // * and when data are buffered, whether to get them off of the heap or to // * transfer them among nodes. // * // * @return // */ // IChunkedIterator chunkedIterator(); }





© 2015 - 2025 Weber Informatics LLC | Privacy Policy