simplifications in DHT Distribution class and more documentation

pull/1/head
Michael Peter Christen 13 years ago
parent e57bf2ca39
commit 5683162bd3

@ -124,6 +124,8 @@ public class ASCII implements Comparator<String> {
}
public final static byte[] getBytes(final String s) {
assert s != null;
assert s.length() < 3 || s.charAt(2) != '@';
int count = s.length();
final byte[] b = new byte[count];
for (int i = 0; i < count; i++) {

@ -30,11 +30,7 @@ import net.yacy.cora.order.Base64Order;
* vertical: scale with number of references for every word
* The vertical scaling is selected using the corresponding reference hash, the url hash
* This has the effect that every vertical position accumulates references for the same url
* and the urls are not spread over all positions of the DHT. To use this effect, the
* horizontal DHT position must be normed to a 'rest' value of a partition size.
* @param wordHash, the hash of the RWI
* @param urlHash, the hash of a reference
* @return a double in the range 0 .. 1.0 (including 0, excluding 1.0), the DHT position
* and the urls are not spread over all positions of the DHT.
*/
public class Distribution {
@ -50,36 +46,70 @@ public class Distribution {
*/
public Distribution(int verticalPartitionExponent) {
assert verticalPartitionExponent > 0;
// the partition exponent is the number of bits that we use for the partition
this.verticalPartitionExponent = verticalPartitionExponent;
this.partitionCount = 1 << verticalPartitionExponent;
// number of partitions that is possible for the given number of partition exponent bits
this.partitionCount = 1 << this.verticalPartitionExponent;
// we use Long.SIZE - 1 as bitlength since we use only the 63 bits of 0..Long.MAX_VALUE
this.shiftLength = Long.SIZE - 1 - this.verticalPartitionExponent;
// the partition size is the cardinal number of possible hash positions for each segment of the DHT
this.partitionSize = 1L << this.shiftLength;
this.partitionMask = (1L << shiftLength) - 1L;
// the partition mask is a bitmask for each partition
this.partitionMask = this.partitionSize - 1L;
}
public int verticalPartitions() {
return 1 << verticalPartitionExponent;
return this.partitionCount;
}
/**
* the horizontal DHT position uses simply the ordering on hashes, the base 64 order to assign a cardinal
* in the range of 0..Long.MAX_VALUE to the word.
* @param wordHash
* @return
*/
public final static long horizontalDHTPosition(byte[] wordHash) {
assert wordHash != null;
assert wordHash[2] != '@';
return Base64Order.enhancedCoder.cardinal(wordHash);
}
/**
* the horizontal DHT distance is the cardinal number between the cardinal position of the hashes of two objects in the DHT
* Since the DHT is closed at the end, a cardinal at the high-end of 0..Long.MAX_VALUE can be very close to a low cardinal number.
* @param from the start DHT position as word hash
* @param to the end DHT position as word hash
* @return the distance of two positions. The maximal distance is Long.MAX_VALUE / 2
*/
public final static long horizontalDHTDistance(final byte[] from, final byte[] to) {
// the dht distance is a positive value between 0 and 1
// if the distance is small, the word more probably belongs to the peer
assert to != null;
assert from != null;
final long toPos = horizontalDHTPosition(to);
final long fromPos = horizontalDHTPosition(from);
return horizontalDHTDistance(fromPos, toPos);
}
/**
* the horizontalDHTDistance computes the closed-at-the-end ordering of two cardinal DHT positions
* @param fromPos the start DHT position as cardinal of the word hash
* @param toPos the end DHT position as cardinal of the word hash
* @return the distance of two positions. The maximal distance is Long.MAX_VALUE / 2
*/
public final static long horizontalDHTDistance(final long fromPos, final long toPos) {
return (toPos >= fromPos) ? toPos - fromPos : (Long.MAX_VALUE - fromPos) + toPos + 1;
}
/**
* the reverse function to horizontalDHTPosition
* This is a bit fuzzy since the horizontalDHTPosition cannot represent all 72 bits of the word hash (Yes, its a HASH!)
* @param l the cardinal position in the DHT
* @return the abstract/computed word of the cardinal.
*/
public final static byte[] positionToHash(final long l) {
// transform the position of a peer position into a close peer hash
byte[] h = Base64Order.enhancedCoder.uncardinal(l);
@ -87,107 +117,69 @@ public class Distribution {
return h;
}
/**
* the partition size is (Long.MAX + 1) / 2 ** e == 2 ** (63 - e)
* compute the position using a specific fragment of the word hash and the url hash:
* - from the word hash take the 63 - <partitionExponent> lower bits
* - from the url hash take the <partitionExponent> higher bits
* in case that the partitionExpoent is 1, only one bit is taken from the urlHash,
* which means that the partition is in two parts.
* With partitionExponent = 2 it is divided in four parts and so on.
* @param wordHash
* @param urlHash
* @return
*/
public final long verticalDHTPosition(final byte[] wordHash, final String urlHash) {
// this creates 1^^e different positions for the same word hash (according to url hash)
assert wordHash != null;
assert urlHash != null;
if (urlHash == null || verticalPartitionExponent < 1) return Distribution.horizontalDHTPosition(wordHash);
// the partition size is (Long.MAX + 1) / 2 ** e == 2 ** (63 - e)
// compute the position using a specific fragment of the word hash and the url hash:
// - from the word hash take the 63 - <partitionExponent> lower bits
// - from the url hash take the <partitionExponent> higher bits
// in case that the partitionExpoent is 1, only one bit is taken from the urlHash,
// which means that the partition is in two parts.
// With partitionExponent = 2 it is divided in four parts and so on.
return (Distribution.horizontalDHTPosition(wordHash) & partitionMask) | (Distribution.horizontalDHTPosition(ASCII.getBytes(urlHash)) & ~partitionMask);
}
/**
* compute a vertical DHT position for a given word
* This is used when a word is searched and the peers holding the word must be computed
* @param wordHash, the hash of the word
* @param verticalPosition (0 <= verticalPosition < verticalPartitions())
* @return a number that can represents a position and can be computed to a word hash again
*/
public final long verticalDHTPosition(final byte[] wordHash, final int verticalPosition) {
assert wordHash != null;
assert wordHash[2] != '@';
assert verticalPosition >= 0 && verticalPosition < verticalPartitions();
long verticalMask = ((long) verticalPosition) << this.shiftLength; // don't remove the cast! it will become an integer result which is wrong.
return (Distribution.horizontalDHTPosition(wordHash) & partitionMask) | verticalMask;
}
public final int verticalDHTPosition(final byte[] urlHash) {
assert urlHash != null;
return (int) (Distribution.horizontalDHTPosition(urlHash) >> this.shiftLength); // take only the top-<partitionExponent> bits
}
/**
* compute all vertical DHT positions for a given word
* This is used when a word is searched and the peers holding the word must be computed
* @param wordHash, the hash of the word
* @param partitions, the number of partitions of the DHT
* @return a vector of long values, the possible DHT positions
* compute the vertical position of a url hash. Thats the same value as second parameter in verticalDHTPosition/2
* @param urlHash
* @return a number from 0..verticalPartitions()
*/
public final long[] verticalDHTPositions(final byte[] wordHash) {
assert wordHash != null;
long[] l = new long[this.partitionCount];
l[0] = Distribution.horizontalDHTPosition(wordHash) & (partitionSize - 1L); // this is the lowest possible position
for (int i = 1; i < this.partitionCount; i++) {
l[i] = l[i - 1] + partitionSize; // no overflow, because we started with the lowest
}
return l;
}
/*
public static void main(String[] args) {
long c1 = Base64Order.enhancedCoder.cardinal("AAAAAAAAAAAA".getBytes());
System.out.println(ASCII.String(Base64Order.enhancedCoder.uncardinal(c1)));
long c2 = Base64Order.enhancedCoder.cardinal("____________".getBytes());
System.out.println(ASCII.String(Base64Order.enhancedCoder.uncardinal(c2)));
Random r = new Random(System.currentTimeMillis());
for (int i = 0; i < 10000; i++) {
long l = r.nextLong();
byte[] h = positionToHash(l);
if (l != Base64Order.enhancedCoder.cardinal(h)) System.out.println(l);
}
public final int verticalDHTPosition(final byte[] urlHash) {
return (int) (Distribution.horizontalDHTPosition(urlHash) >> this.shiftLength); // take only the top-<partitionExponent> bits
}
*/
public static void main(String[] args) {
// java -classpath classes de.anomic.yacy.yacySeed hHJBztzcFn76
// java -classpath classes de.anomic.yacy.yacySeed hHJBztzcFG76 M8hgtrHG6g12 3
// test the DHT position calculation
String wordHash = "hHJBztzcFn76";
//double dhtd;
byte[] wordHash = UTF8.getBytes("hHJBztzcFn76");
long dhtl;
int partitionExponent = 0;
Distribution partition = new Distribution(0);
if (args.length == 3) {
// the horizontal and vertical position calculation
String urlHash = args[1];
partitionExponent = Integer.parseInt(args[2]);
dhtl = partition.verticalDHTPosition(UTF8.getBytes(wordHash), urlHash);
} else {
// only a horizontal position calculation
dhtl = Distribution.horizontalDHTPosition(UTF8.getBytes(wordHash));
}
//System.out.println("DHT Double = " + dhtd);
int partitionExponent = 4;
Distribution partition = new Distribution(partitionExponent);
dhtl = Distribution.horizontalDHTPosition(wordHash);
System.out.println("DHT Long = " + dhtl);
System.out.println("DHT as Double from Long = " + ((double) dhtl) / ((double) Long.MAX_VALUE));
//System.out.println("DHT as Long from Double = " + (long) (Long.MAX_VALUE * dhtd));
//System.out.println("DHT as b64 from Double = " + positionToHash(dhtd));
System.out.println("DHT as b64 from Long = " + ASCII.String(Distribution.positionToHash(dhtl)));
System.out.print("all " + (1 << partitionExponent) + " DHT positions from doubles: ");
/*
double[] d = dhtPositionsDouble(wordHash, partitionExponent);
for (int i = 0; i < d.length; i++) {
if (i > 0) System.out.print(", ");
System.out.print(positionToHash(d[i]));
}
System.out.println();
*/
System.out.print("all " + (1 << partitionExponent) + " DHT positions from long : ");
long[] l = partition.verticalDHTPositions(UTF8.getBytes(wordHash));
for (int i = 0; i < l.length; i++) {
if (i > 0) System.out.print(", ");
System.out.print(ASCII.String(Distribution.positionToHash(l[i])));
System.out.println("all " + partition.verticalPartitions() + " DHT positions from long : ");
for (int i = 0; i < partition.verticalPartitions(); i++) {
long l = partition.verticalDHTPosition(wordHash, i);
System.out.println(ASCII.String(Distribution.positionToHash(l)));
}
System.out.println();
long c1 = Base64Order.enhancedCoder.cardinal("AAAAAAAAAAAA".getBytes());
System.out.println(ASCII.String(Base64Order.enhancedCoder.uncardinal(c1)));
long c2 = Base64Order.enhancedCoder.cardinal("____________".getBytes());
System.out.println(ASCII.String(Base64Order.enhancedCoder.uncardinal(c2)));
}
}

@ -197,9 +197,9 @@ public class DHTSelection {
int redundancy,
Map<String, Seed> regularSeeds) {
// this method is called from the search target computation
final long[] dhtVerticalTargets = seedDB.scheme.verticalDHTPositions(wordhash);
Seed seed;
for (long dhtVerticalTarget : dhtVerticalTargets) {
for (int verticalPosition = 0; verticalPosition < seedDB.scheme.verticalPartitions(); verticalPosition++) {
long dhtVerticalTarget = seedDB.scheme.verticalDHTPosition(wordhash, verticalPosition);
wordhash = Distribution.positionToHash(dhtVerticalTarget);
Iterator<Seed> dhtEnum = getAcceptRemoteIndexSeeds(seedDB, wordhash, redundancy, false);
int c = Math.min(seedDB.sizeConnected(), redundancy);

@ -273,6 +273,7 @@ public class Dispatcher {
* then no additional IO is necessary.
*/
private void enqueueContainersToCloud(final List<ReferenceContainer<WordReference>>[] containers) {
assert (containers.length == this.seeds.scheme.verticalPartitions());
if (this.transmissionCloud == null) return;
ReferenceContainer<WordReference> lastContainer;
byte[] primaryTarget;

@ -161,9 +161,10 @@ public class NetworkGraph {
final Iterator<byte[]> i = query.query_include_hashes.iterator();
eventPicture.setColor(RasterPlotter.GREY);
while (i.hasNext()) {
final long[] positions = seedDB.scheme.verticalDHTPositions(i.next());
for (final long position : positions) {
angle = cyc + (360.0d * ((position) / DOUBLE_LONG_MAX_VALUE));
byte[] wordHash = i.next();
for (int verticalPosition = 0; verticalPosition < seedDB.scheme.verticalPartitions(); verticalPosition++) {
long position = seedDB.scheme.verticalDHTPosition(wordHash, verticalPosition);
angle = cyc + (360.0d * (position / DOUBLE_LONG_MAX_VALUE));
eventPicture.arcLine(cx, cy, cr - 20, cr, angle, true, null, null, -1, -1, -1, false);
}
}

@ -147,14 +147,14 @@ public final class TemplateEngine {
private final static byte[] ul = "_".getBytes();
private final static byte[] alternative_which = ASCII.getBytes(" type=\"alternative\" which=\"");
private final static byte[] multi_num = ASCII.getBytes(" type=\"multi\" num=\"");
private final static byte[] open_endtag = ASCII.getBytes("</");
private final static byte[] close_quotetagn = ASCII.getBytes("\">\n");
private final static byte[] close_tagn = ASCII.getBytes(">\n");
private final static byte[] PP = ASCII.getBytes("%%");
private final static byte[] hash_brackopen_slash = ASCII.getBytes("#(/");
private final static byte[] brackclose_hash = ASCII.getBytes(")#");
private final static byte[] alternative_which = " type=\"alternative\" which=\"".getBytes();
private final static byte[] multi_num = " type=\"multi\" num=\"".getBytes();
private final static byte[] open_endtag = "</".getBytes();
private final static byte[] close_quotetagn = "\">\n".getBytes();
private final static byte[] close_tagn = ">\n".getBytes();
private final static byte[] PP = "%%".getBytes();
private final static byte[] hash_brackopen_slash = "#(/".getBytes();
private final static byte[] brackclose_hash = ")#".getBytes();
/**

Loading…
Cancel
Save