我在尝试从另一个 csv 文件开始创建 csv 文件时遇到性能问题。 这是原始文件的样子:
country,state,co,olt,olu,splitter,ont,cpe,cpe.latitude,cpe.longitude,cpe.customer_class,cpe.phone,cpe.ip,cpe.subscriber_id
COUNTRY-0001,STATE-0001,CO-0001,OLT-0001,OLU0001,SPLITTER-0001,ONT-0001,CPE-0001,28.21487,77.451775,ALL,SIP:+674100002743@IMS.COMCAST.NET,SIP:E28EDADA06B2@IMS.COMCAST.NET,CPE_SUBSCRIBER_ID-QHLHW4
COUNTRY-0001,STATE-0002,CO-0002,OLT-0002,OLU0002,SPLITTER-0002,ONT-0002,CPE-0002,28.294018,77.068924,ALL,SIP:+796107443092@IMS.COMCAST.NET,SIP:58DD999D6466@IMS.COMCAST.NET,CPE_SUBSCRIBER_ID-AH8NJQ
可能有数百万行这样的行,我已经检测到 1.280.000 行的问题。
这是算法:
File csvInputFile = new File(csv_path);
int blockSize = 409600;
brCsvInputFile = new BufferedReader(frCsvInputFile, blockSize);
String line = null;
StringBuilder sbIntermediate = new StringBuilder();
skipFirstLine(brCsvInputFile);
while ((line = brCsvInputFile.readLine()) != null) {
createIntermediateStringBuffer(sbIntermediate, line.split(REGEX_COMMA));
}
private static void skipFirstLine(BufferedReader br) throws IOException {
String line = br.readLine();
String[] splitLine = line.split(REGEX_COMMA);
LOGGER.debug("First line detected! ");
createIndex(splitLine);
createIntermediateIndex(splitLine);
}
private static void createIndex(String[] splitLine) {
LOGGER.debug("START method createIndex.");
for (int i = 0; i < splitLine.length; i++)
headerIndex.put(splitLine[i], i);
printMap(headerIndex);
LOGGER.debug("COMPLETED method createIndex.");
}
private static void createIntermediateIndex(String[] splitLine) {
LOGGER.debug("START method createIntermediateIndex.");
com.tekcomms.c2d.xml.model.v2.Metadata_element[] metadata_element = null;
String[] servicePath = newTopology.getElement().getEntity().getService_path().getLevel();
if (newTopology.getElement().getMetadata() != null)
metadata_element = newTopology.getElement().getMetadata().getMetadata_element();
LOGGER.debug(servicePath.toString());
LOGGER.debug(metadata_element.toString());
headerIntermediateIndex.clear();
int indexIntermediateId = 0;
for (int i = 0; i < servicePath.length; i++) {
String level = servicePath[i];
LOGGER.debug("level is: " + level);
headerIntermediateIndex.put(level, indexIntermediateId);
indexIntermediateId++;
// its identificator is going to be located to the next one
headerIntermediateIndex.put(level + "ID", indexIntermediateId);
indexIntermediateId++;
}
// adding cpe.latitude,cpe.longitude,cpe.customer_class, it could be
// better if it would be metadata as well.
String labelLatitude = newTopology.getElement().getEntity().getLatitude();
// indexIntermediateId++;
headerIntermediateIndex.put(labelLatitude, indexIntermediateId);
String labelLongitude = newTopology.getElement().getEntity().getLongitude();
indexIntermediateId++;
headerIntermediateIndex.put(labelLongitude, indexIntermediateId);
String labelCustomerClass = newTopology.getElement().getCustomer_class();
indexIntermediateId++;
headerIntermediateIndex.put(labelCustomerClass, indexIntermediateId);
// adding metadata
// cpe.phone,cpe.ip,cpe.subscriber_id,cpe.vendor,cpe.model,cpe.customer_status,cpe.contact_telephone,cpe.address,
// cpe.city,cpe.state,cpe.zip,cpe.bootfile,cpe.software_version,cpe.hardware_version
// now i need to iterate over each Metadata_element belonging to
// topology.element.metadata
// are there any metadata?
if (metadata_element != null && metadata_element.length != 0)
for (int j = 0; j < metadata_element.length; j++) {
String label = metadata_element[j].getLabel();
label = label.toLowerCase();
LOGGER.debug(" ==label: " + label + " index_pos: " + j);
indexIntermediateId++;
headerIntermediateIndex.put(label, indexIntermediateId);
}
printMap(headerIntermediateIndex);
LOGGER.debug("COMPLETED method createIntermediateIndex.");
}
读取整个数据集,1.280.000 行需要 800 毫秒!所以问题出在这个方法上
private static void createIntermediateStringBuffer(StringBuilder sbIntermediate, String[] splitLine) throws ClassCastException,
NullPointerException {
LOGGER.debug("START method createIntermediateStringBuffer.");
long start, end;
start = System.currentTimeMillis();
ArrayList<String> hashes = new ArrayList<String>();
com.tekcomms.c2d.xml.model.v2.Metadata_element[] metadata_element = null;
String[] servicePath = newTopology.getElement().getEntity().getService_path().getLevel();
LOGGER.debug(servicePath.toString());
if (newTopology.getElement().getMetadata() != null) {
metadata_element = newTopology.getElement().getMetadata().getMetadata_element();
LOGGER.debug(metadata_element.toString());
}
for (int i = 0; i < servicePath.length; i++) {
String level = servicePath[i];
LOGGER.debug("level is: " + level);
if (splitLine.length > getPositionFromIndex(level)) {
String name = splitLine[getPositionFromIndex(level)];
sbIntermediate.append(name);
hashes.add(name);
sbIntermediate.append(REGEX_COMMA).append(HashUtils.calculateHash(hashes)).append(REGEX_COMMA);
LOGGER.debug(" ==sbIntermediate: " + sbIntermediate.toString());
}
}
// end=System.currentTimeMillis();
// LOGGER.info("COMPLETED adding name hash. " + (end - start) + " ms. " + (end - start) / 1000 + " seg.");
// adding cpe.latitude,cpe.longitude,cpe.customer_class, it should be
// better if it would be metadata as well.
String labelLatitude = newTopology.getElement().getEntity().getLatitude();
if (splitLine.length > getPositionFromIndex(labelLatitude)) {
String lat = splitLine[getPositionFromIndex(labelLatitude)];
sbIntermediate.append(lat).append(REGEX_COMMA);
}
String labelLongitude = newTopology.getElement().getEntity().getLongitude();
if (splitLine.length > getPositionFromIndex(labelLongitude)) {
String lon = splitLine[getPositionFromIndex(labelLongitude)];
sbIntermediate.append(lon).append(REGEX_COMMA);
}
String labelCustomerClass = newTopology.getElement().getCustomer_class();
if (splitLine.length > getPositionFromIndex(labelCustomerClass)) {
String customerClass = splitLine[getPositionFromIndex(labelCustomerClass)];
sbIntermediate.append(customerClass).append(REGEX_COMMA);
}
// end=System.currentTimeMillis();
// LOGGER.info("COMPLETED adding lat,lon,customer. " + (end - start) + " ms. " + (end - start) / 1000 + " seg.");
// watch out metadata are optional, it can appear as a void chain!
if (metadata_element != null && metadata_element.length != 0)
for (int j = 0; j < metadata_element.length; j++) {
String label = metadata_element[j].getLabel();
LOGGER.debug(" ==label: " + label + " index_pos: " + j);
if (splitLine.length > getPositionFromIndex(label)) {
String actualValue = splitLine[getPositionFromIndex(label)];
if (!"".equals(actualValue))
sbIntermediate.append(actualValue).append(REGEX_COMMA);
else
sbIntermediate.append("").append(REGEX_COMMA);
} else
sbIntermediate.append("").append(REGEX_COMMA);
LOGGER.debug(" ==sbIntermediate: " + sbIntermediate.toString());
}//for
sbIntermediate.append("\n");
end = System.currentTimeMillis();
LOGGER.info("COMPLETED method createIntermediateStringBuffer. " + (end - start) + " ms. ");
}
如您所见,此方法将预先计算的行添加到 StringBuffer,从输入 csv 文件读取每一行,从这些行计算新数据,最后将生成的行添加到 StringBuffer,所以最后我可以使用以下命令创建文件该缓冲区。
我已经运行了jconsole,我可以看到没有内存泄漏,我可以看到代表对象创建的锯齿和GC重新收集garbaje。它永远不会超越内存堆阈值。
我注意到的一件事是,向 StringBuffer 添加新行所需的时间是在很少的毫秒范围内完成的(5,6,10),但随着时间的推移而增加,达到(100-200)女士和我怀疑在不久的将来会有更多,所以这可能是战马。
我尝试分析代码,我知道有3个for循环,但它们很短,第一个循环仅迭代8个元素:
for (int i = 0; i < servicePath.length; i++) {
String level = servicePath[i];
LOGGER.debug("level is: " + level);
if (splitLine.length > getPositionFromIndex(level)) {
String name = splitLine[getPositionFromIndex(level)];
sbIntermediate.append(name);
hashes.add(name);
sbIntermediate.append(REGEX_COMMA).append(HashUtils.calculateHash(hashes)).append(REGEX_COMMA);
LOGGER.debug(" ==sbIntermediate: " + sbIntermediate.toString());
}
}
我测量了从分割线获取名称所需的时间,它毫无值(value),0毫秒,与calculateHash方法相同,0毫秒。
另一个循环实际上是相同的,迭代 0 到 n,其中 n 是一个非常小的整数,例如 3 到 10,所以我不明白为什么需要更多时间来完成该方法,唯一的事情我发现向缓冲区添加新行的过程变得很慢。
我正在考虑一种生产者消费者多线程策略,一个读取器线程读取每一行并将它们放入循环缓冲区中,另一个线程逐个读取它,处理它们并将预先计算的行添加到 StringBuffer 中,这是线程安全,当文件被完全读取时,读取线程向另一个线程发送一条消息,告诉它们停止。最后我必须将此缓冲区保存到文件中。你怎么认为?这是一个好主意?
最佳答案
I am thinking about a producer consumer multi threaded strategy, a reader thread that reads every line and put them into a circular buffer, another threads take it one by one, process them and add a precalculated line to the StringBuffer, which is thread safe, when the file is fully readed, the reader thread sends a message to to the another threads telling them to stop. Finally i have to save this buffer to a file. What do you think? this is a good idea?
也许吧,但是工作量很大,我会先尝试一些更简单的方法。
line.split(REGEX_COMMA)
您的REGEX_COMMA
是一个被编译成正则表达式一百万次的字符串。这很简单,但我会尝试使用 Pattern
相反。
你的 split 产生了很多垃圾。也许您应该通过手动将输入拆分为重复使用的 ArrayList<String>
来避免这种情况。 (只有几行)。
如果您需要的只是将结果写入文件,那么最好避免构建一个巨大的字符串。也许是List<String>
甚至是List<StringBuilder>
会更好,也许直接写入缓冲流可以。
您似乎只使用 ASCII。您的编码取决于平台,这可能意味着您使用的是 UTF-8,这可能很慢。切换到更简单的编码可能会有所帮助。
与 byte[]
合作而不是String
很可能会有所帮助。字节大小是字符的一半,读取文件时无需转换。您所做的所有操作都可以同样轻松地使用字节完成。
One thing i have noticed is that the time needed for add a new line to the StringBuffer is completed within a very few ms range, (5,6,10), but is raising with time, to (100-200) ms and i suspect more in a near future, so probably this is the battle horse.
这就是调整大小,可以通过使用建议的 ArrayList<String>
来加快速度。 ,因为要复制的数据量要少得多。当缓冲区变大时写出数据也可以。
I have meassured the time needed to get the name from the splitline and it is worthless, 0 ms, the same to calculateHash method, 0 ms.
切勿使用currentTimeMillis
为此 nanoTime
严格来说更好。使用分析器。分析器的问题在于它改变了它应该测量的内容。作为穷人的分析器,您可以计算可疑方法内部花费的所有时间的总和,并将其与总时间进行比较。
运行程序时CPU负载是多少以及GC做了什么?
关于java - 在 Java 中创建大型 csv 文件变得非常慢,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/27295356/