JDK-5072161 : OutOfMemoryError while using GZIPOutputStream
  • Type: Bug
  • Status: Closed
  • Resolution: Duplicate
  • Component: core-libs
  • Sub-Component: java.util.jar
  • Priority: P2
  • Affected Version: 1.3.1_12
  • OS: generic
  • CPU: generic
  • Submit Date: 2004-07-08
  • Updated Date: 2006-03-20
  • Resolved Date: 2006-03-20
Related Reports
Duplicate :  
Relates :  
Relates :  
Relates :  
Relates :  
Relates :  
Relates :  
Description
Customer is using his J2EE Server with JDK 1.3.1_12. The SAP J2EE Engine is running the Enterprise Portal and serving compressed output. The java.util.zip.GZIPOutputStream which is used calls into the OpenSource compression library zlib (www.gzip.org/zlib) which we ship (even in Tiger!) in version 1.1.3 from 9 July 1998. After some hours the server crashes with an OOM.

- The latest zlib is: 1.2.1, from 17. Nov. 2003, check the ChangeLog there are MANY fixes.
- An alternative usage of a pure-java compression makes the crash disappear.
- SAP swaped the zlib-sources in the JDK ( just to narrow this issue down!) to the latest one ( ... and build their own VM ): No OOM anymore.
###@###.### 10/16/04 01:02 GMT

I've changed this bug from
Synopsis: OutOfMemoryError while using GZIPOutputStream, zlib memleak
to
(ref) Finalizer thread sometimes can't keep up with other threads

to more accurately reflect the underlying problem.
###@###.### 2005-1-10 23:30:21 GMT

Comments
EVALUATION In the prior 2 weeks, feedback (to me via private email) was that this bug can be closed. Since the underlying issue still seems to persist (e.g. 4797189 and 6293787) I'm marking it a duplicate.
2006-03-20

EVALUATION Since this bug has been in the incomplete state for a year, has also received no updates during that time, and there are other equally descriptive bugs (e.g. 4797189, 6293787), this will be closed in 2 weeks unless further information is provided that can move it out ouf the incomplete/need-more-info state.
2006-03-03

EVALUATION Upgrading zlib to 1.2.1 is a good idea, for a post-Tiger release. The zlib shipped with the JDK is not a plain 1.1.3; bug fixes have been applied, so the upgrade is non-trivial. We are not aware of any memory leaks in the jar/zip code relating to our use of zlib, despite efforts in the past to find such leaks. We need a small reproducible test case, preferably one that can be reproduced using the latest Tiger bits. ###@###.### 2004-07-09 Here is a test program based on a JDC comment by "hubt": ---------------------------------------------------------------- import java.io.*; import java.util.zip.*; public class Bug { public static void main(String args[]) throws Throwable { boolean gc = args.length > 0 && args[0].equals("gc"); boolean close = args.length > 0 && args[0].equals("close"); byte[] bytes = "Hello World".getBytes(); int i = 0; ByteArrayOutputStream os = new ByteArrayOutputStream(1000); for(i = 0; i < 1000 * 1000 ; i++ ) { os.reset(); GZIPOutputStream gos = new GZIPOutputStream(os); gos.write(bytes,0,bytes.length); gos.finish(); if (close) gos.close(); if (i % 1000 == 0) { if (gc) { System.gc(); System.runFinalization(); } System.out.println("Iterations: " + i); } } } } ---------------------------------------------------------------- The program fails when called with no arguments, while succeeding when called with either the "gc" or "close" argument, showing that this is very likely an instance of... "I think we're seeing the usual problem with Java's inability to collect non-heap-memory resources in a timely fashion. The finalizable objects which can unmap the memory will be collected, but the GC is not aware of the urgency, since they appear to be simple small Java objects." In the absence of a general solution to the non-heap resource exhaustion problem, users of classes with close() methods should make sure to call them as soon as possible. The problem is more thoroughly explained in 5092131: using 1 MB pagesize throws "not enough space" error with 32bit JVM We don't know whether this is the original submitter's problem, but I suspect that this is the root cause of all "Zip out of memory" problems in Tiger. ###@###.### 2004-09-04 To begin with, I produced a smaller version of the customer's ZipedContentGenerator.java (from the attachments) --------------------------------------------------------------------- import java.io.ByteArrayOutputStream; import java.util.Random; import java.util.zip.DeflaterOutputStream; public class ZipedContentGenerator { public static void main(String[] args) { for (int i=0; i<20; i++) new CompressingThread().start(); } } class CompressingThread extends Thread { final private static String OUTPUT_CHARS = "AaBbCcDdEeFfGgHhIiJjKkLlMmNnOoPpQqRrSsTtUuVvWwXxYyZz"; public void run() { while (true) { ByteArrayOutputStream baos = new ByteArrayOutputStream(); DeflaterOutputStream dos = new DeflaterOutputStream(baos); try { dos.write(generateContent().getBytes()); dos.close(); dos = null; baos = null; // ATTENTION !!!!!!!!!!!!!!!!!!! // IF you run the finalizers here, the Problem goes away: // System.runFinalization(); } catch (Exception e) { e.printStackTrace(); } } } private String generateContent() { StringBuffer result = new StringBuffer(100000); addRandomOutput(result, 100000, 5); return result.toString(); } public void addRandomOutput(StringBuffer output, int size, int iWsPaddingRatio) { Random rnd = new Random(); char key; char newChar; int ratioCounter = 0; for (int i = 0; i < size; i++) { if (ratioCounter == iWsPaddingRatio) { newChar = OUTPUT_CHARS.charAt(rnd.nextInt(52)); ratioCounter = 0; } else { newChar = ' '; ratioCounter ++; } output.append(newChar); } } } --------------------------------------------------------------------- (Even though the code correctly calls the close() method...) this causes OutOfMemoryErrors when run, for example, using java -Xmx32m ZipedContentGenerator on solaris-sparc. Then I reduced this to eliminate any complex classes, using a class with a simple finalize method: --------------------------------------------------------------------- public class OOM { public static void main(String[] args) { for (int i=0; i<20; i++) new Thread() { public void run() { for (int j=0; ; j++) { new HasFinalizer(); //System.runFinalization(); }}}.start(); } } class HasFinalizer { private static int i = 0; protected void finalize() { i++; } } --------------------------------------------------------------------- This causes an OOME in the same way. The OOME can be eliminated by commenting out the finalize() method or by uncommenting the call to runFinalization(). So... this is a generic GC problem. There are 20 threads doing nothing but generating finalizable objects, and only one thread running the finalizers. No wonder it can't keep up. I'll transfer this to the GC team for comment. ###@###.### 2004-12-20 19:43:33 GMT This is not a GC issue. We're not going to fix this bug by changing the finalizer infrastructure. Doing so risks introducing all sorts of nasty performance and scalability issues. (We're certainly not going to redesign the finalizer infrastructure in an update release in order to fix an escalated bug!) Finalizers are fundamentally broken; the best we can do is to encourage people not to rely upon them. This has been our standing policy for many years. Are you absolutely sure that there isn't a memory leak in our copy of zlib? How do you know? Have you tried substituting a newer version of zlib, as SAP did, and if so did that help? If user code is correctly closing deflaters/inflaters/zip*streams using try/finally blocks then finalization should not even be an issue. If it is an issue then perhaps we're doing something wrong in our Java code, and somehow keeping a heap object that refers to native memory alive longer than actually necessary. There are well-known techniques for dealing with such problems. I'm moving this bug back to classes_util_jarzip. ###@###.### 2005-1-11 04:01:05 GMT This is not a finalization issue at all, at least not for the given test code. The test runs out of Java heap memory because it spends most of its time with the garbage collector disabled. The Deflater class passes its input and output buffers directly to zlib after acquiring them via the GetPrimitiveArrayCritical JNI call. In HotSpot this effectively disables collection for all threads until the corresponding ReleasePrimitiveArrayCritical calls are made. When a thread that's not in one of these critical sections tries to allocate heap memory while collection is disabled it receives an OutOfMemoryError. Most of the time in this program is spent in these critical sections in the native zlib deflateBytes method, hence the time until the first OutOfMemoryError is thrown is essentially linear in the amount of memory allocated by the VM at startup. Invoking the System.runFinalization method at the end of each iteration in each thread appears to fix the problem only because it introduces some noise into the system and allows collections to happen more frequently. The same OOMEs are thrown even if the finalize method in the Deflater class is removed. If the test code is an accurate reflection of what's going on in production then I'm afraid the options are limited. There's an open RFE (6186200) that would change the behavior of HotSpot in this situation so that a thread would wait until a collection can be attempted rather than throw an OOME when the GC is disabled due to another thread being in a critical section. That change is, however, highly unlikely to be suitable for a 1.3.1 update release. The best workaround I've come up with is to cause each thread to sleep a little bit after each write to the ByteArrayOutputStream in order to allow other threads a chance to invoke the collector if they need to. This can be done by wrapping a fairly simple filtering stream around the ByteArrayOutputStream, e.g., static class DallyingOutputStream extends OutputStream { private OutputStream dst; public DallyingOutputStream(OutputStream dst) { this.dst = dst; } public void write(int b) throws IOException { dst.write(b); } public void write(byte[] b, int off, int len) throws IOException { try { Thread.currentThread().sleep(50); } catch (InterruptedException x) { } dst.write(b, off, len); } public void flush() throws IOException { dst.flush(); } public void close() throws IOException { dst.close(); } } ... ByteArrayOutputStream baos = new ByteArrayOutputStream(); DallyingOutputStream dyos = new DallyingOutputStream(baos); DeflaterOutputStream dos = new DeflaterOutputStream(dyos); ... With this change I've been able to run the test program for several minutes without failure in a 64MB heap. This workaround reduces throughput, obviously, but at least it avoids the OutOfMemoryErrors. At this point I'm wondering if the test code really is an accurate reflection of what's going on in production. Answers to the following questions would be helpful: (1) What is the actual OutOfMemoryError, and stack trace, seen in production? (2) Does the production code close each DeflaterOutputStream when finished? (3) Does the production code actually convert and compress 100k characters at a time? (4) Does the production code accumulate output in a ByteArrayOutputStream, as in the test, or is the output written to, say, a network connection? (5) Does the production code do all this with this many threads (20), or more, or less? On how many processors does it run? On what OS? (6) Exactly which DeflaterOutputStream constructor does the production code use, and with what arguments? ###@###.### 2005-1-24 15:51:51 GMT While looking into this problem two bugs in the DeflaterOutputStream and Deflater classes were discovered; see 6223075 and 6223076. Fixing these bugs appears to help the customer's test case run for a longer time, but still not indefinitely. Nevertheless it might be worthwhile to send a test build to the customer to see if fixing these two bugs helps in production. ###@###.### 2005-1-28 20:27:35 GMT Customer answered the six questions thous: (1) What is the actual OutOfMemoryError, and stack trace, seen in production? Please see the attachment (Q1.ZIP) you will find the logs from a production environment and some screen shots of the consoles just after the out of memory error was reproduced. Note: When the screen shots were made the server was running with JDK 1.3.1.11 (2) Does the production code close each DeflaterOutputStream when finished? Yes, it does. (3) Does the production code actually convert and compress 100k characters at a time? Yes, it does. (4) Does the production code accumulate output in a ByteArrayOutputStream, as in the test, or is the output written to, say, a network connection? To be absolutely precise the web container uses its GzipResponseStream, which extends javax.servlet.ServletOutputStream, which extends java.io.OutputStream The inheritance in the provided test-case is similar. java.io.ByteArrayOutputStream extends java.io.OutputStream (5) Does the production code do all this with this many threads (20), or more, or less? On how many processors does it run? On what OS? Here I devided your questions in three sub-questions: (5.1) Does the production code do all this with this many threads (20), or more, or less? By default the application threads (each one could serve one request) are set to 40. This means that on heavy load the threads could become 40 (if the default value is not changed). (5.2) On how many processors does it run? SAP J2EE Application Server 6.20 runs only on JDK 1.3.1. One application server node could scale, if it uses from 1 to 4 CPUs. There are no customer cases where we have more application server nodes than CPUs. (5.3) On what OS? The detailed list of supported OS by the J2EE Engine 6.20 could be found in the PAM portal follow this link https://websmp105.sap-ag.de/pam then from the left menu choose consequently 1. SAP NetWeaver 2. SAP NetWeaver Componnets (< SAP NW 04) 3. SAP WebAS\SAP WEB AS 6.20 then from the tab menu on the left choose JSE Platforms. A list of all suported OS will appear. The described configuration in Answer 5.1 could be set to any of the operation systems. However the reported problem about the issue discussed in this CSN was observed only on Windows and HPUX platforms. (6) Exactly which DeflaterOutputStream constructor does the production code use, and with what arguments? In the production code we use an instance of java.util.zip.GZIPOutputStream, which extends java.util.zip.DeflaterOutputStream, which extends java.io.FilterOutputStream Here is a part of the production code showing the instantiation of the zipping output stream (FilterOutputStream) private void instantiateStream() throws IOException { if (SBasic.gzipImpConstructor == null) { gzipstream = new GZIPOutputStream(servletoutput); } else { // initialize the server to use an external zipping library } } gzipstream instance is of class java.io.FilterOutputStream Actually what you are interested in is the code in the if statement (the one in else statement is executed in case the server is going to use an external zipping library). Consequently the answer to your question about the constructor is: FilterOutputStream gzipstream = new GZIPOutputStream(servletoutput); the parameter is a ServletOutputStream ###@###.### 2005-2-04 11:42:36 GMT The above answers confirm that the supplied test code is not an accurate reflection of the customer's application. In particular the customer's code uses GZIPOutputStream while the test code uses DeflaterOutputStream. This may be a critical difference. It turns out that the GZIPOutputStream class in 1.3.1 has the same problem as the DeflaterOutputStream class in that release, namely that invoking the close() method will not cause the underlying deflater's end() method to be invoked since the private usesDefaultDeflater field is never set. I've updated 6223075 accordingly. I suggest sending a new build to the customer containing fixes for both 6223075 and 6223076 to see if this makes a difference in production. ###@###.### 2005-2-07 04:51:19 GMT ----- Tested a 1.3.1xx based binary containing fixes for CR#{6223075 and 6223076} on both dual and single cpu machines, using the stand alone test case supplied by Cu. The testing took place for more than 48Hrs. No OOM error is observed. I've used the same command line options that were used by Cu. Here is partial GC output: /home2/rai/java/1.3.1-14/build/solaris-sparc/bin/java -Xms1024m -Xmx1024m -XX:SurvivorRatio=2 -XX:TargetSurvivorRatio=90 -XX:NewRatio=6 -verbose:gc ZipedContentGenerator [GC 1158K->1083K(1047936K), 0.0059966 secs] [GC 5981K->4805K(1047936K), 0.0038203 secs] ............................................. ............................................. [GC 1043797K->1042921K(1047936K), 0.1272579 secs] [GC 1044177K->1043667K(1047936K), 0.1291067 secs] [GC 1044941K->1044419K(1047936K), 0.1716337 secs] [GC 1045659K->1045035K(1047936K), 0.0235835 secs] [Full GC 1046484K->9207K(1047936K), 9.4994934 secs] [GC 10406K->9964K(1047936K), 0.0544921 secs] [GC 11534K->10689K(1047936K), 0.0254829 secs] [GC 11811K->11240K(1047936K), 0.0819868 secs] What is observed is all the minor collections resulted in very minimal memory retrieval and one major [Full GC] collection retrieved good amount of memory. Ran the same test case with everything same except the jdk binary which does not has the fixes. And OOM error is reproduced. ###@###.### 2005-2-12 00:27:28 GMT
2005-02-12

SUGGESTED FIX use a newer|latest zlib
2004-09-06

PUBLIC COMMENTS -
2004-09-06