lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "wangzhijiang999" <wangzhijiang...@aliyun.com>
Subject 答复:答复:mmap confusion in lucene
Date Tue, 15 Jul 2014 09:09:45 GMT
Hi 308181687,
 
         I also tested in this way. If print every byte, the OS cache will consume
the size of file at last,about 800M.
 
for (int j = 0; j < len; ++j){        System.out.println(buff.get());}    

 
 
If just call buff.get() in loop, the OS cache will consume only 8M at last.
for (int j = 0; j < len; ++j){        byte b=buff.get();}    
 
The buff.get() means reading the byte at this buffer's current position, and then increments
the position. But actually if you do not use the value from buff.get(), FS  will not read
the disk. And I monitored the disk read and cache condition by dstat -md command to confirm
that the disk read will not increase for the second test.
As you said, the jvm is so smart that if you do not use the data  , it will not read from
disk. As my previous understanding, as long as you use get method to fetch data, it should
read from disk no matter whether you actually use the data or not. I will continue researching
on it to find the real reason. 
       



------------------------------------------------------------------发件人:308181687 <308181687@qq.com>发送时间:2014年7月15日(星期二)
13:04收件人:java-user <java-user@lucene.apache.org>主 题:Re:答复:mmap
confusion in luceneHi, Zhiiang It seems that the jvm is smart enough to ignore the unused
code. Try the following code:RandomAccessFile raf = new RandomAccessFile(new File("/root/xx.txt"),
"r");FileChannel rafc = raf.getChannel();ByteBuffer buff = rafc.map(FileChannel.MapMode.READ_ONLY,
0, rafc.size());int len=buff.limit();byte b = 0;for (int i = 0; i < len; i++){b + = buff.get();}The
java process will consume the expected 800M share memory. But if change the line of " b +
= buff.get()" to "b = buff.get()", the java process will not consume so much share memory,
i guess that the jvm is smart enough to directly skip to the the last pos of the bytebuffer
.Thanks & Best Regards!‍------------------ Original ------------------From: "java-user@lucene.apache.org
wan";<wangzhijiang999@aliyun.com>;Date: Tue, Jul 15, 2014 10:44 AMTo: "java-user"<java-user@lucene.apache.org>;
Subject: 答复:mmap confusion in luceneHi Uwe,Thank you for always help. For my first testing
I am clear of it, it is becuase the OS cache the whole file because of copying data to java
heap and it does not free the page, then I see 800M used by cache in the end.But for my last
two testings, the OS has freed all the previous cached pages, so I see the cache used only
4M in the end.Maybe I am not very clear of the internal kernel mechanism. As I understand,
the kernel will swap out the page when the memory resource is limited or the cached page is
not used for long time. The first condition is not satisfied in my testing, because the OS
still has 30G memory available for use. For the second condition, although the bytes are copied
to java heap in first test, but when the program ends to quit, the OS still reserve the cache.
In the last test, the OS released the page even in the running process of program. Would you
give me some further explaination for this? I am very appreciated.Zhiiang Wang------------------------------------------------------------------发件人:Uwe
Schindler <uwe@thetaphi.de>发送时间:2014年7月14日(星期一) 18:13收件人:java-user
<java-user@lucene.apache.org>; wangzhijiang999 <wangzhijiang999@aliyun.com>主 题:RE:
mmap confusion in luceneThis is very easy to explain:In the first part you copy the whole
memory mapped stuff into a on-heap byte array. You allocate this byte array in total and you
then do a copy (actually this is a standard libc copy call) of the whole file. To do this
copy, the underlying OS will need to swap in the whole file, because it "sees" that you want
to read the whole file anyway (because of the size of they copy operation).The other example
reads the stuff byte by byte in a Java for-loop. The operating system has no idea how to optimize
that, so whenever you cross page boundaries it will swap in another buffer. Because of internal
kernel-page-garbage collection, the pages swapped in are freed much faster. This is OS specific.In
general copying a random access file to java heap with mmap is just the wrong use case. Lucene
never does this! The idea behind mmap is to *not copy* the data and work on the mmapped region
directly (using random access). The OS cache logic will then use statistics about which pages
were actually used and keep them longer in FS cache than those used one time and then no longer
used for very long time.Uwe-----Uwe SchindlerH.-H.-Meier-Allee 63, D-28213 Bremenhttp://www.thetaphi.deeMail:
uwe@thetaphi.de> -----Original Message-----> From: wangzhijiang999 [mailto:wangzhijiang999@aliyun.com]>
Sent: Monday, July 14, 2014 11:58 AM> To: java-user> Subject: mmap confusion in lucene>
> Hi everybody, I found a problem confused me when I tested the mmap> feature in lucene.
I tested to read a file size of 800M by mmap method like> below:> > RandomAccessFile
raf = new RandomAccessFile(new File(path), "r");> FileChannel rafc = raf.getChannel();ByteBuffer
buff => rafc.map(FileChannel.MapMode.READ_ONLY, 0, rafc.size());> int len=buff.limit();
byte[] b = new byte[len]; for (int i = 0; i < len;> i++){ b[i] = buff.get(); }> After
the program finished, the linux cache will be consumed about 800M.> > > RandomAccessFile
raf = new RandomAccessFile(new File(path), "r");> FileChannel rafc = raf.getChannel();ByteBuffer
buff => rafc.map(FileChannel.MapMode.READ_ONLY, 0, rafc.size());> int len=buff.limit();
for (int i = 0; i < len; i++){ Byte b= buff.get(); }> But in this way, the linux cache
will be consumed just 4M.> > > RandomAccessFile raf = new RandomAccessFile(new File(path),
"r");> FileChannel rafc = raf.getChannel();ByteBuffer buff => rafc.map(FileChannel.MapMode.READ_ONLY,
0, rafc.size());> int len=buff.limit(); byte[] b = new byte[len]; for (int i = 0; i <
len;> i++){ b[i] = buff.get();> b[i]=0; }> In this way, the linux cache will be also
consumed 4M.> > The whole content of the file should be read for above three tests,
but for> the last two testings, the linux system only cached 4M .> Would somebody give
me the explaination about this? Thanks in advane.> > Zhijiang Wang> ---------------------------------------------------------------------To
unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.orgFor additional commands, e-mail:
java-user-help@lucene.apache.org
Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message