We Keep Coding

Java code runs out of space memory on AWS but not MacOSX

I need another set of eyes on this.
I've written out a zip file into hundreds of gigabytes with this exact code with no modifications locally on MacOSX.
With 100% unchanged code, just deployed to an AWS instance running Ubuntu, this same code runs into Out of Memory issues (heap space).
Here's the code that's being run, streaming MyBatis to a CSV file on disk:
File directory = new File(feedDirectory);
File file;
try {
file = File.createTempFile(("feed-" + providerCode + "-"), ".csv", directory);
} catch (IOException e) {
throw new RuntimeException("Unable to create file to write feed to disk: " + e.getMessage(), e);
}
String filePath = file.getAbsolutePath();
log.info(String.format("File name for %s feed is %s", providerCode, filePath));
// output file
try (FileOutputStream out = new FileOutputStream(file)) {
streamData(out, providerCode, startDate, endDate);
} catch (IOException e) {
throw new RuntimeException("Unable to write feed to file: " + e.getMessage());
}
public void streamData(OutputStream outputStream, String providerCode, Date startDate, Date endDate) throws IOException {
try (CSVPrinter printer = CsvUtil.openPrinter(outputStream)) {
StreamingHandler<FStay> handler = stayPrintingHandler(printer);
warehouse.doForAllStaysByProvider(providerCode, startDate, endDate, handler);
}
}
private StreamingHandler<FStay> stayPrintingHandler(CSVPrinter printer) {
StreamingHandler<FStay> handler = new StreamingHandler<>();
handler.setHandler((stay) -> {
try {
EXPORTER.writeStay(printer, stay);
} catch (IOException e) {
log.error("Issue with writing output: " + e.getMessage(), e);
}
});
return handler;
}
// The EXPORTER method
import org.apache.commons.csv.CSVPrinter;
public void writeStay(CSVPrinter printer, FStay stay) throws IOException {
List<Object> list = asList(stay);
printer.printRecord(list);
}
List<Object> asList(FStay stay) {
List<Object> list = new ArrayList<>(46);
list.add(stay.getUid());
list.add(stay.getProviderCode());
//....
return list;
}
Here's a graph of the JVM heap space (using jvisualvm) when I run this locally. I've run this consistently with of Java 8 (jdk1.8.0_51 and 1.8.0_112) locally and have gotten great results. Even written out a terabyte of data.
^ In the above, the max heap space is set to 4 gigs, and the most it ever increases to is 1.5 gigs, before going back down to around 500 MB, while streaming data to the CSV file as it's supposed to.
However, when I run this on Ubuntu with jdk 1.8.0_111, the exact same operation will not complete, running out of heap space (java.lang.OutOfMemoryError: Java heap space)
I've upped the Xmx value from 8 gigs to 16 to 25 gigs, and still run out of heap space. Meanwhile... the total size of the file is only 10 Gigs in total... which really perplexes me.
Here's what the JVisualVm graph looks like on the Ubuntu box:
I've no doubt it's the exact same code running in both environments, with the same operation being performed in each (same database server providing the same data)
The only differences I can think of at this point are:
Operating system - Ubuntu vs Mac OS X
Hosted VM in AWS vs hard metal laptop
Network speed is faster in AWS between database and Ubuntu server
JDK version is 1.8.0_111 in Ubuntu, tried 1.8.0_51 and 1.8.0_112 locally
Can anyone help shed any light on this problem?
Update
I've tried replacing all the 'try-with-resources' statements with explicit flush/close statements and no luck.
What's more, I tried to force a garbage collection on the Ubuntu box as soon as I started to see the data come in, and it had no effect-- there is something definitely stopping the heap from being collected on the Ubuntu machine... while running the exact same code on OS X let me write the full enchilada again no problem.
Update 2
In addition to the differences in the environments above, the only other difference I can think of is if the connection between the servers in AWS is so fast that it streams the data faster than it can flush the data to disk... but that still doesn't explain the issue where I only have 10 gigs of data total, and it blows up a JVM with 20 Gigs of heap space.
Is there any likelihood of there being a bug at the Ubuntu/Java level for this?
Update 3
Tried replacing the output of the CSVPrinter to use an entirely separate library (OpenCSV's CSVWriter in lieu of Apache's CSV library) and the same result occurs.
As soon as this code starts receiving data from the database, the heap starts blowing up and the garbage collector fails to reclaim any memory... but only on Ubuntu. On OS X, everything is reclaimed immediately and the heap never grows.
I've also tried flushing the stream after every write, but had no luck with that as well.
Update 4
Got the heap dump to print out, and according to this I should be looking at the database driver. Specifically the InboundDataHandler in amazon's redshift driver.
I'm using myBatis with a custom result handler. I tried setting the result handler to effectively do nothing when it gets a result (new ResultHandler<>() { // method overridden to do literally nothing}) and I know I'm not holding on to any references there.
Since it's the InboundDataHandler defined by AWS/Redshift... it makes me think it may be lower than the myBatis level... either:
Error in the SqlSessionFactory I'm setting up
Bug in the Redshift driver that only pops up in Ubuntu / AWS
Bug in the result handler I have overwritten
Here's the heap dump screenshot:
Here's where I'm setting up my SqlSessionFactoryBean:
#Bean
public javax.sql.DataSource redshiftDataSource() throws ClassNotFoundException {
log.info("Got to datasource config");
// Dynamically load driver at runtime.
Class.forName(dataWarehouseDriver);
DataSource dataSource = new DataSource();
dataSource.setURL(dataWarehouseUrl);
dataSource.setUserID(dataWarehouseUsername);
dataSource.setPassword(dataWarehousePassword);
return dataSource;
}
#Bean
public SqlSessionFactoryBean sqlSessionFactory() throws ClassNotFoundException {
SqlSessionFactoryBean factoryBean = new SqlSessionFactoryBean();
factoryBean.setDataSource(redshiftDataSource());
return factoryBean;
}
Here's the myBatis code I'm running as a test to verify that it's not me holding on to records in my ResultHandler:
warehouse.doForAllStaysByProvider(providerCode, startDate, endDate, new ResultHandler<FStay>() {
#Override
public void handleResult(ResultContext<? extends FStay> resultContext) {
// do nothing
}
});
Is there a way I can force the SQL connection to not hang on to records or something? I'll again re-iterate that on my local machine, there is no issue with this memory leak... it only surfaces when running the code in the hosted AWS environment. And in both cases, the Database driver and server are the same.
Update 6
I think it's finally fixed. Thanks to all who pointed me in the direction of the heap dump. That helped narrow it down to the offending class in a huge way.
After that, I did some research on the AWS redshift driver, and it explicitly says that your clients should specify a limit for any operations on large data. So I found out how to do that in my myBatis configuration:
<select id="doForAllStaysByProvider" fetchSize="1000" resultMap="FStayResultMap">
select distinct
f_stay.uid,
And this did the trick.
Mind you, this isn't necessary even when handling much larger data sets downloaded remotely from AWS (Database in AWS, code executing on laptop at home), and this shouldn't be necessary since I'm overriding the myBatis ResultHandler<> which handles each row individually and never holds on to any objects.
Yet something funky happens with the AWS redshift jdbc driver only when it's run in AWS (database in aws, code executing in AWS instance) which causes this InboundDataHandler to never release its resources, unless a fetchSize is specified.
Here's the heap of the server running now, getting much further than it ever has before in AWS, with the heap space never moving above 500Mb, and after i hit 'force gc' in jvisualvm, it shows the 'used' heap at less than 100mb:
Thanks again in a huge way to all those who helped guide this!

Finally figured out a solution.
The heap dump was the biggest aid-- it indicated the InboundDataHandler class of Amazon's RedShift/postgres JDCB driver was the prime culprit.
The code to set up the SqlSession appeared legit, so traveling over to Amazon's documentation landed this gem:
To avoid client-side out-of-memory errors when retrieving large data
sets using JDBC, you can enable your client to fetch data in batches
by setting the JDBC fetch size parameter.
We hadn't run into this before, as we stream results with custom ResultHandlers in MyBatis... but there seems to be something different when the AWS Redshift JDBC driver is running on AWS itself vs outside AWS connecting in.
Taking the guidance from the documentation, we added a 'fetchSize' to our MyBatis select query:
<select id="doForAllStaysByProvider" fetchSize="1000" resultMap="FStayResultMap">
select distinct
f_stay.uid,
And voila! Everything worked swimmingly. This is the only change we made and the heap never went above a couple hundred MBs.
You can see in one of the above graphs where the heap goes off the charts, as soon as the data started to be received on Amazon, the heap marches right up linearly and never reclaims an ounce of heap space once it starts.
My guess is the Redshift JDBC driver is doing something different when it's in Amazon's environment for some kind of optimization... that's all I can think of to explain the behavior.
Clearly Amazon knows what's going on since they documented it up front. I may not know the full 'why' of what's happening, but at least everything is resolved in what appears to be a satisfactory way.
Thanks to all those who helped.

Java Memory Error on Importing .xlsx files into R [duplicate]

The xlsx package can be used to read and write Excel spreadsheets from R. Unfortunately, even for moderately large spreadsheets, java.lang.OutOfMemoryError can occur. In particular,
Error in .jcall("RJavaTools", "Ljava/lang/Object;", "invokeMethod", cl, :
java.lang.OutOfMemoryError: Java heap space
Error in .jcall("RJavaTools", "Ljava/lang/Object;", "newInstance", .jfindClass(class), :
java.lang.OutOfMemoryError: GC overhead limit exceeded
(Other related exceptions are also possible but rarer.)
A similar question was asked regarding this error when reading spreadsheets.
Importing a big xlsx file into R?
The main advantage of using Excel spreadsheets as a data storage medium over CSV is that you can store multiple sheets in the same file, so here we consider a list of data frames to be written one data frame per worksheet. This example dataset contains 40 data frames, each with two columns of up to 200k rows. It is designed to be big enough to be problematic, but you can change the size by altering n_sheets and n_rows.
library(xlsx)
set.seed(19790801)
n_sheets <- 40
the_data <- replicate(
n_sheets,
{
n_rows <- sample(2e5, 1)
data.frame(
x = runif(n_rows),
y = sample(letters, n_rows, replace = TRUE)
)
},
simplify = FALSE
)
names(the_data) <- paste("Sheet", seq_len(n_sheets))
The natural method of writing this to file is to create a workbook using createWorkbook, then loop over each data frame calling createSheet and addDataFrame. Finally the workbook can be written to file using saveWorkbook. I've added messages to the loop to make it easier to see where it falls over.
wb <- createWorkbook()
for(i in seq_along(the_data))
{
message("Creating sheet", i)
sheet <- createSheet(wb, sheetName = names(the_data)[i])
message("Adding data frame", i)
addDataFrame(the_data[[i]], sheet)
}
saveWorkbook(wb, "test.xlsx")
Running this in 64-bit on a machine with 8GB RAM, it throws the GC overhead limit exceeded error while running addDataFrame for the first time.
How do I write large datasets to Excel spreadsheets using xlsx?

This is a known issue:
http://code.google.com/p/rexcel/issues/detail?id=33
While unresolved, the issue page links to a solution by Gabor Grothendieck suggesting that the heap size should be increased by setting the java.parameters option before the rJava package is loaded. (rJava is a dependency of xlsx.)
options(java.parameters = "-Xmx1000m")
The value 1000 is the number of megabytes of RAM to allow for the Java heap; it can be replaced with any value you like. My experiments with this suggest that bigger values are better, and you can happily use your full RAM entitlement. For example, I got the best results using:
options(java.parameters = "-Xmx8000m")
on the machine with 8GB RAM.
A further improvement can be obtained by requesting a garbage collection in each iteration of the loop. As noted by #gjabel, R garbage collection can be performed using gc(). We can define a Java garbage collection function that calls the Java System.gc() method:
jgc <- function()
{
.jcall("java/lang/System", method = "gc")
}
Then the loop can be updated to:
for(i in seq_along(the_data))
{
gc()
jgc()
message("Creating sheet", i)
sheet <- createSheet(wb, sheetName = names(the_data)[i])
message("Adding data frame", i)
addDataFrame(the_data[[i]], sheet)
}
With both these code fixes, the code ran as far as i = 29 before throwing an error.
One technique that I tried unsuccessfully was to use write.xlsx2 to write the contents to file at each iteration. This was slower than the other code, and it fell over on the 10th iteration (but at least part of the contents were written to file).
for(i in seq_along(the_data))
{
message("Writing sheet", i)
write.xlsx2(
the_data[[i]],
"test.xlsx",
sheetName = names(the_data)[i],
append = i > 1
)
}

Building on #richie-cotton answer, I found adding gc() to the jgc function kept the CPU usage low.
jgc <- function()
{
gc()
.jcall("java/lang/System", method = "gc")
}
My previous for loop still struggled with the original jgc function, but with extra command, I no longer run into GC overhead limit exceeded error message.

Solution for the above error:
Please use the below mentioned r - code:
detach(package:xlsx)
detach(package:XLConnect)
library(openxlsx)
And, try to import the file again and you will not get any error as it works for me.

Restart R and, before loading the R packages, insert:
options(java.parameters = "-Xmx2048m")
or
options(java.parameters = "-Xmx8000m")

You can also use gc() inside the loop if you are writing row by row. gc() stands for garbage collection. gc() can be used in any case of memory issue.

I was having issues with write.xlsx() rather than reading.... but then realised that I had accidentally been running 32bit R. Swapping it out to 64bit has fixed the issue.

Encog - EarlyStoppingStrategy with validation set

I would like to stop training a network once I see the error calculated from the validation set starts to increase. I'm using a BasicNetwork with RPROP as the training algorithm, and I have the following training iteration:
double validationError = 999.999;
while(!stop){
train.iteration(); //weights are updated here
System.out.println("Epoch #" + epoch + " Error : " + train.getError()) ;
//I'm just comparing to see if the error on the validation set increased or not
if (network.calculateError(validationSet) < validationError)
validationError = network.calculateError(validationSet);
else
//once the error increases I stop the training.
stop = true ;
System.out.println("Epoch #" + epoch + "Validation Error" + network.calculateError(validationSet));
epoch++;
}
train.finishTraining();
Obviously this isn't working because the weights have already been changed before figuring out if I need to stop training or not. Is there anyway I can take a step back and use the old weights?
I also see the EarlyStoppingStrategy class which is probably what I need to use by using the addStrategy() method. However, I really don't understand why the EarlyStoppingStrategy constructor takes both the validation set and the test set. I thought it would only need the validation set and the test set shouldn't be used at all until I test the output of the network.

Encog's EarlyStoppingStrategy class implements an early stopping strategy according to this paper:
Proben1 | A Set of Neural Network Benchmark Problems and Benchmarking Rules
(a full cite is included in the Javadoc)
If you just want to stop as soon as a validation set's error no longer improves you may just want to use the Encog SimpleEarlyStoppingStrategy, found here:
org.encog.ml.train.strategy.end.SimpleEarlyStoppingStrategy
Note, that SimpleEarlyStoppingStrategy requires Encog 3.3.

Running Calabash XML From Code

I downloaded Calabash XML a couple of days back and got it working easily enough from the command prompt. I then tried to run it from Java code I noticed there was no API (e.g. the Calabash main method is massive with code calls to everywhere). To get it working was very messy as I had to copy huge chunks from the main method to a wrapper class, and divert from the System.out to a byte array output stream (and eventually into a String) i.e.
...
ByteArrayOutputStream baos = new ByteArrayOutputStream (); // declare at top
...
WritableDocument wd = null;
if (uri != null) {
URI furi = new URI(uri);
String filename = furi.getPath();
FileOutputStream outfile = new FileOutputStream(filename);
wd = new WritableDocument(runtime,filename,serial,outfile);
} else {
wd = new WritableDocument(runtime,uri,serial, baos); // new "baos" parameter
}
The performance seems really, really slow e.g. i ran a simple filter 1000 times ...
<p:filter>
<p:with-option name="select" select="'/result/meta-data/neighbors/document/title'" />
</p:filter>
On average each time took 17ms which doesn't seem like much but my spring REST controller with calls to Mongo DB and encryption calls etc take on average 3/4 ms.
Has anyone encountered this when running Calabash from code? Is there something I can do to speed things up?
For example, I this is being called each time -
XProcRuntime runtime = new XProcRuntime(config);
Can this be created once and reused? Any help is appreciated as I don't want to have to pay money to use Calamet but really want to get Xproc working from code to an acceptable performance.

For examples on how you could integrate XMLCalabash in a framework, I can mention Servlex by Florent Georges. You'd have to browse the code to find the relevant bit, but last time I looked it shouldn't be too hard to find:
http://servlex.net/
XMLCalabash wasn't build for speed unfortunately. I am sure that if you run profile, and can find some hotspots, Norm Walsh would be interested to hear about it.
Alternative is to look into Quixprox, which is derived from XMLCalabash:
https://code.google.com/p/quixproc/
I am also very sure that if you can send Norm a patch to improve the main class for better integration, he'd be interested to hear about it. In fact, the code should be on github, just fork it, fix it, and do a pull request..
HTH!

Java application performs very slow (10-100 times slower than on Windows, Linux, AIX)

I need your help about performance problems running our corporate Java application on HP\UX server. Application is standalone tool which synchronizes data over several data bases into one, communicates with remote control on XML-RPC protocol and uses local Derby (Java DB) data base instance to hold configuration data etc. We don not have performance problems on other environments on the same load like Windows XP, Linux and AIX which use Sun JVM. After series of test we found out that most time consuming was communication with Derby data base. Most time is spent on reading from socket and this time is greater in 10-100 times than on other platforms. We know for sure that Derby works fine, we’ve got CPU reserve (usage is about 30%-40%), so most probable reason is transport layer between local data base and application.
Is there a way to diagnose socket I\O problems on HP-UX or maybe there is some possible limitations that can be configured? Maybe there is necessary JVM option? Any ideas from your side would be highly appreciated.
We’ve tried to optimize JVM options according to http://publib.boulder.ibm.com/infocenter/wasinfo/v6r1/index.jsp?topic=/com.ibm.websphere.wsfep.multiplatform.doc/info/ae/ae/tprf_tunejvm_v61.html but didn’t get any significant improvement.
JVM info:
Java HotSpot(TM) 64-Bit Server VM (19.1-b02-jinteg:2011mar11-16:46 PA2.0W (aCC_AP), mixed mode)
Java: version 1.6.0.10, vendor "Hewlett-Packard Company"
We use following instance:
OS: HP-UX (B.11.23)
Architecture: PA_RISC2.0W 64bit
Processors: 2
Total physical memory size: 4 088 MB
Swap size: 4 090 MB
Here is example of slow running code. It takes several seconds to execute on HP while on Windows it takes 10-30ms:
/** Template to communicate with local db. */
SimpleJdbcTemplate jdbcTemplate;
#Transactional(readOnly = true)
public List<JobLogEntry> getLastLogs(Integer dbnr, JobDataType dtype) {
try {
String uid = jdbcTemplate.queryForObject("SELECT session_uuid FROM "
+ tableName + " WHERE id=(SELECT max(id) FROM "
+ tableName + " WHERE dbnr=? AND dtype=?)",
String.class, dbnr, dtype.name());
List<JobLogEntry> list = jdbcTemplate.query("SELECT id, dbnr, dtype, zeit, level, message FROM "
+ tableName
+ " WHERE dbnr=? AND dtype=? AND session_uuid=? ORDER BY ID",
new ConRowMapper(), dbnr, dtype.name(), uid);
return list;
} catch (org.springframework.dao.EmptyResultDataAccessException e) {
return new ArrayList<JobLogEntry>();
}
}
class ConRowMapper implements RowMapper<JobLogEntry> {
private final SimpleDateFormat dateFormat = new SimpleDateFormat("yyyy.MM.dd HH:mm:ss");
/**
* Maps rows.
*/
public JobLogEntry mapRow(ResultSet rs, int rowNum) throws SQLException {
return new JobLogEntry(rs.getInt("dbnr"),
rs.getString("dtype"),
dateFormat.format(rs.getTimestamp("zeit")),
rs.getString("level"),
rs.getString("message"));
}
}
Thanks in advance for all your ideas

I wonder about the method getLastLogs(). Why query to get the session UUID and then turn around and use it in another query? I would guess that it's possible to do it in one query.
When you say Derby, it makes me think that only Java accesses that database. Is that true? Do you know that it's optimized well (e.g. proper indexes for every WHERE clause)?
Do you use connection pooling? That way you can pay the cost of creating connections up front and amortize it over all the queries you run.
I see jdbcTemplate, so you must be using Spring. I'd get the debug or trace interceptor wired in and see where the time is being spent.
I'd also recommend Visual VM 1.3.2 will all the plugins installed. It will give you a lot more data.

Probable reason could be slow and blockong GC work on HP-UX. Try removing redundant System.gc() calls and use some JVM GC options to otimize :)
See nice presentation about HP performance tuning: http://www.scribd.com/doc/47433278/Javamemorymanagemen