我会构建文件,以便它可以就地使用。即不以这种方式加载。由于您有可变长度的记录,您可以构建每个记录位置的数组,然后按顺序放置键,以便您可以对数据执行二进制搜索。(或者您可以使用自定义哈希表)然后您可以使用隐藏数据实际存储在文件中而不是转换为数据对象的事实的方法包装它。
如果你做了所有这些,“加载”阶段就变得多余了,你就不需要创建这么多的对象了。
这是一个很长的例子,但希望能展示什么是可能的。
import vanilla.java.chronicle.Chronicle;
import vanilla.java.chronicle.Excerpt;
import vanilla.java.chronicle.impl.IndexedChronicle;
import vanilla.java.chronicle.tools.ChronicleTest;
import java.io.IOException;
import java.util.*;
public class Main {
static final String TMP = System.getProperty("java.io.tmpdir");
public static void main(String... args) throws IOException {
String baseName = TMP + "/test";
String[] keys = generateAndSave(baseName, 100 * 1000 * 1000);
long start = System.nanoTime();
SavedSortedMap map = new SavedSortedMap(baseName);
for (int i = 0; i < keys.length / 100; i++) {
long l = map.lookup(keys[i]);
// System.out.println(keys[i] + ": " + l);
}
map.close();
long time = System.nanoTime() - start;
System.out.printf("Load of %,d records and lookup of %,d keys took %.3f seconds%n",
keys.length, keys.length / 100, time / 1e9);
}
static SortedMap<String, Long> generateMap(int keys) {
SortedMap<String, Long> ret = new TreeMap<>();
while (ret.size() < keys) {
long n = ret.size();
String key = Long.toString(n);
while (key.length() < 9)
key = '0' + key;
ret.put(key, n);
}
return ret;
}
static void saveData(SortedMap<String, Long> map, String baseName) throws IOException {
Chronicle chronicle = new IndexedChronicle(baseName);
Excerpt excerpt = chronicle.createExcerpt();
for (Map.Entry<String, Long> entry : map.entrySet()) {
excerpt.startExcerpt(2 + entry.getKey().length() + 8);
excerpt.writeUTF(entry.getKey());
excerpt.writeLong(entry.getValue());
excerpt.finish();
}
chronicle.close();
}
static class SavedSortedMap {
final Chronicle chronicle;
final Excerpt excerpt;
final String midKey;
final long size;
SavedSortedMap(String baseName) throws IOException {
chronicle = new IndexedChronicle(baseName);
excerpt = chronicle.createExcerpt();
size = chronicle.size();
excerpt.index(size / 2);
midKey = excerpt.readUTF();
}
// find exact match or take the value after.
public long lookup(CharSequence key) {
if (compareTo(key, midKey) < 0)
return lookup0(0, size / 2, key);
return lookup0(size / 2, size, key);
}
private final StringBuilder tmp = new StringBuilder();
private long lookup0(long from, long to, CharSequence key) {
long mid = (from + to) >>> 1;
excerpt.index(mid);
tmp.setLength(0);
excerpt.readUTF(tmp);
if (to - from <= 1)
return excerpt.readLong();
int cmp = compareTo(key, tmp);
if (cmp < 0)
return lookup0(from, mid, key);
if (cmp > 0)
return lookup0(mid, to, key);
return excerpt.readLong();
}
public static int compareTo(CharSequence a, CharSequence b) {
int lim = Math.min(a.length(), b.length());
for (int k = 0; k < lim; k++) {
char c1 = a.charAt(k);
char c2 = b.charAt(k);
if (c1 != c2)
return c1 - c2;
}
return a.length() - b.length();
}
public void close() {
chronicle.close();
}
}
private static String[] generateAndSave(String baseName, int keyCount) throws IOException {
SortedMap<String, Long> map = generateMap(keyCount);
saveData(map, baseName);
ChronicleTest.deleteOnExit(baseName);
String[] keys = map.keySet().toArray(new String[map.size()]);
Collections.shuffle(Arrays.asList(keys));
return keys;
}
}
生成 2 GB 的原始数据并执行一百万次查找。它以这样一种方式编写,即加载和查找使用很少的堆。( << 1 MB )
ls -l /tmp/test*
-rw-rw---- 1 peter peter 2013265920 Dec 11 13:23 /tmp/test.data
-rw-rw---- 1 peter peter 805306368 Dec 11 13:23 /tmp/test.index
/tmp/test created.
/tmp/test, size=100000000
Load of 100,000,000 records and lookup of 1,000,000 keys took 10.945 seconds
每次查找使用哈希表查找会更快,因为它是 O(1) 而不是 O(ln N),但实现起来更复杂。