Details

    • Type: Bug Bug
    • Status: Closed
    • Priority: Major Major
    • Resolution: Fixed
    • Affects Version/s: None
    • Fix Version/s: 3.1, 4.0-ALPHA
    • Component/s: core/store
    • Labels:
      None
    • Lucene Fields:
      New, Patch Available

      Description

      There are a couple of issues with it:

      1. FileChannel.transferFrom documents that it may not copy the number of bytes requested, however we don't check the return value. So need to fix the code to read in a loop until all bytes were copied..
      2. When calling addIndexes() w/ very large segments (few hundred MBs in size), I ran into the following exception (Java 1.6 – Java 1.5's exception was cryptic):
        Exception in thread "main" java.io.IOException: Map failed
            at sun.nio.ch.FileChannelImpl.map(FileChannelImpl.java:770)
            at sun.nio.ch.FileChannelImpl.transferToTrustedChannel(FileChannelImpl.java:450)
            at sun.nio.ch.FileChannelImpl.transferTo(FileChannelImpl.java:523)
            at org.apache.lucene.store.FSDirectory.copy(FSDirectory.java:450)
            at org.apache.lucene.index.IndexWriter.addIndexes(IndexWriter.java:3019)
        Caused by: java.lang.OutOfMemoryError: Map failed
            at sun.nio.ch.FileChannelImpl.map0(Native Method)
            at sun.nio.ch.FileChannelImpl.map(FileChannelImpl.java:767)
            ... 7 more
        

      I changed the impl to something like this:

      long numWritten = 0;
      long numToWrite = input.size();
      long bufSize = 1 << 26;
      while (numWritten < numToWrite) {
        numWritten += output.transferFrom(input, numWritten, bufSize);
      }
      

      And the code successfully adds the indexes. This code uses chunks of 64MB, however that might be too large for some applications, so we definitely need a smaller one. The question is how small so that performance won't be affected, and it'd be great if we can let it be configurable, however since that API is called by other API, such as addIndexes, not sure it's easily controllable.

      Also, I read somewhere (can't remember now where) that on Linux the native impl is better and does copy in chunks. So perhaps we should make a Linux specific impl?

      1. FileCopyTest.java
        4 kB
        Shai Erera
      2. LUCENE-2537.patch
        10 kB
        Shai Erera
      3. LUCENE-2537.patch
        2 kB
        Shai Erera

        Activity

        Hide
        Shai Erera added a comment - - edited

        Oh .. found the thread we discussed that on the list, to which I've actually last posted w/ the following text:

        I've Googled around a bit and came across this: http://markmail.org/message/l67bierbmmedrfw5. Apparently, there's a long standing bug against SUN since May 2006 (http://bugs.sun.com/view_bug.do?bug_id=6431344) that's still open and reports the exact same behavior that I'm seeing.

        If I understand correctly, this might be a Windows limitation and is expected to work well on Linux. I'll give it a try. But this makes me think if we should keep the current behavior for Linux-based directories, and fallback to the chunks approach for Windows ones? Since eventually I'll be running on Linux, I don't want to lose performance ...

        This isn't the first that we've witnessed the "write once, run everywhere" misconception of Java . I'm thinking if in general we should have a Windows/Linux FSDirectory impl, or handlers, to prepare for future cases as well. Mike already started this with LUCENE-2500 (DirectIOLinuxDirectory). Instead of writing a Directory, perhaps we could have a handler object or something, or a generic LinuxDirectory that impls some stuff the 'linux' way. In FSDirectory we already have code which detects the OS and JRE used to decide between Simple, NIO and MMAP Directories ...

        Show
        Shai Erera added a comment - - edited Oh .. found the thread we discussed that on the list, to which I've actually last posted w/ the following text: I've Googled around a bit and came across this: http://markmail.org/message/l67bierbmmedrfw5 . Apparently, there's a long standing bug against SUN since May 2006 ( http://bugs.sun.com/view_bug.do?bug_id=6431344 ) that's still open and reports the exact same behavior that I'm seeing. If I understand correctly, this might be a Windows limitation and is expected to work well on Linux. I'll give it a try. But this makes me think if we should keep the current behavior for Linux-based directories, and fallback to the chunks approach for Windows ones? Since eventually I'll be running on Linux, I don't want to lose performance ... This isn't the first that we've witnessed the "write once, run everywhere" misconception of Java . I'm thinking if in general we should have a Windows/Linux FSDirectory impl, or handlers, to prepare for future cases as well. Mike already started this with LUCENE-2500 (DirectIOLinuxDirectory). Instead of writing a Directory, perhaps we could have a handler object or something, or a generic LinuxDirectory that impls some stuff the 'linux' way. In FSDirectory we already have code which detects the OS and JRE used to decide between Simple, NIO and MMAP Directories ...
        Hide
        Shai Erera added a comment -

        I wrote a test which compares FileChannel API to intermediate buffer copies. The test runs each method 3 times and reports the best time of each. It can be run w/ different file and chunk sizes.

        Here are the results of copying a 1GB file using different chunk sizes (the chunk is used as the intermediate buffer size as well).

        Machine spec:

        • Linux, 64-bit (IBM) JVM
        • 2xQuad (+hyper-threading) - 16 cores overall
        • 16GB RAM
        • SAS HD
        Chunk Size FileChannel Intermediate Buffer Diff
        64K 1865 1528 -18%
        128K 1660 1526 -9%
        512K 1514 1493 -2%
        1M 1552 2072 +33%
        2M 1488 1559 5%
        4M 1596 1831 13%
        16M 1563 1964 21%
        64M 1494 2442 39%
        128M 1469 2445 40%

        For small buffer sizes, intermediate byte[] copies is preferable. However, FileChannel method performs pretty much consistently, irregardless of the buffer size (except for the first run), while the byte[] approach degrades a lot, as the buffer size increases.

        I think, given these results, we can use the FileChannel method w/ a chunk size of 4 (or even 2) MB, to be on the safe side and don't eat up too much RAM?

        Show
        Shai Erera added a comment - I wrote a test which compares FileChannel API to intermediate buffer copies. The test runs each method 3 times and reports the best time of each. It can be run w/ different file and chunk sizes. Here are the results of copying a 1GB file using different chunk sizes (the chunk is used as the intermediate buffer size as well). Machine spec: Linux, 64-bit (IBM) JVM 2xQuad (+hyper-threading) - 16 cores overall 16GB RAM SAS HD Chunk Size FileChannel Intermediate Buffer Diff 64K 1865 1528 -18% 128K 1660 1526 -9% 512K 1514 1493 -2% 1M 1552 2072 +33% 2M 1488 1559 5% 4M 1596 1831 13% 16M 1563 1964 21% 64M 1494 2442 39% 128M 1469 2445 40% For small buffer sizes, intermediate byte[] copies is preferable. However, FileChannel method performs pretty much consistently, irregardless of the buffer size (except for the first run), while the byte[] approach degrades a lot, as the buffer size increases. I think, given these results, we can use the FileChannel method w/ a chunk size of 4 (or even 2) MB, to be on the safe side and don't eat up too much RAM?
        Hide
        Michael McCandless added a comment -

        Nice results Shai!

        I think, given these results, we can use the FileChannel method w/ a chunk size of 4 (or even 2) MB, to be on the safe side and don't eat up too much RAM?

        +1

        Show
        Michael McCandless added a comment - Nice results Shai! I think, given these results, we can use the FileChannel method w/ a chunk size of 4 (or even 2) MB, to be on the safe side and don't eat up too much RAM? +1
        Hide
        Shai Erera added a comment -

        Patch copies the files in chunks of 2MB. All core tests pass. I'll wait a day or two in case someone wants to suggests a different approach, or chunk size limit before I commit.

        Show
        Shai Erera added a comment - Patch copies the files in chunks of 2MB. All core tests pass. I'll wait a day or two in case someone wants to suggests a different approach, or chunk size limit before I commit.
        Hide
        Shai Erera added a comment -

        Patch adds the following:

        • FSIndexOutput overrides copyBytes to do optimized copies as well (when possible).
        • CompoundFileWriter changed to call copyBytes, instead of using an intermediate buffer.
        • Some minor typos and code corrections.

        NOTE: with the changes to CFW, CheckAbort is accessed only after an entire file is written. The estimated amount of work is the length copied. This means that if abort() is called, it might take a tad longer until CFW will detect it.
        I don't think it's serious, since (1) abort() is not called often, (2) for small segments this will probably have no effect (OneMerge was accessed roughly every ~2MB copied) and (3) as reported, the optimized copy is faster when using FileChannel, therefore the time that passes between checks may not be that long.
        And there's a (4) – for really large segments, the amount of work done to merge them is far larger than copying them into the CFS. Therefore the chances that abort() will be called during that process is relatively small ...

        All tests pass.

        Show
        Shai Erera added a comment - Patch adds the following: FSIndexOutput overrides copyBytes to do optimized copies as well (when possible). CompoundFileWriter changed to call copyBytes, instead of using an intermediate buffer. Some minor typos and code corrections. NOTE: with the changes to CFW, CheckAbort is accessed only after an entire file is written. The estimated amount of work is the length copied. This means that if abort() is called, it might take a tad longer until CFW will detect it. I don't think it's serious, since (1) abort() is not called often, (2) for small segments this will probably have no effect (OneMerge was accessed roughly every ~2MB copied) and (3) as reported, the optimized copy is faster when using FileChannel, therefore the time that passes between checks may not be that long. And there's a (4) – for really large segments, the amount of work done to merge them is far larger than copying them into the CFS. Therefore the chances that abort() will be called during that process is relatively small ... All tests pass.
        Hide
        Shai Erera added a comment -

        If there are no objections, I'll commit this today,

        Show
        Shai Erera added a comment - If there are no objections, I'll commit this today,
        Hide
        Shai Erera added a comment -

        Committed revision 979161 (3x).
        Committed revision 979185.

        Show
        Shai Erera added a comment - Committed revision 979161 (3x). Committed revision 979185.
        Hide
        Grant Ingersoll added a comment -

        Bulk close for 3.1

        Show
        Grant Ingersoll added a comment - Bulk close for 3.1

          People

          • Assignee:
            Shai Erera
            Reporter:
            Shai Erera
          • Votes:
            0 Vote for this issue
            Watchers:
            1 Start watching this issue

            Dates

            • Created:
              Updated:
              Resolved:

              Development