Re: GCJ slower than JDK?

[Bryce McKinlay <bryce at waitaki dot otago dot ac dot nz>]

Fernando Lozano wrote:

>>> Testing on a directory with about 3.000 jpg files and getting about 
>>> 76 duplicates, I found the program is three times slower when 
>>> compiled with gcj -O2 than when compiled and executed bu a Java SDK. 
>>
>>
>> If you'd like to post your small test program, we can take a look at 
>> it and see why it might be slower. 
>
>
> Thanks a lot. Its attached to this message. Please forgiveme my 
> variable names, I am not a native english speaker. :-) 


I tried it out. Without any changes, GCJ came out just slightly ahead of 
JDK 1.4beta2 w/ Hotspot client VM:

$ gcj -O2 DupFinder.java --main=DupFinder
$ /usr/bin/time ./a.out /usr/lib/* > /dev/null
58.32user 1.91system 1:00.27elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1326major+623minor)pagefaults 0swaps

$ /usr/bin/time java DupFinder /usr/lib/* > /dev/null
62.04user 1.86system 1:04.17elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k

But really the problem is with your code. By giving the Adler32 a buffer 
to workwith instead of calling it one byte at a time, it will run much 
faster:

--- DupFinder.java.old  Tue Jan 29 16:17:27 2002
+++ DupFinder.java      Tue Jan 29 16:14:09 2002
@@ -100,6 +100,8 @@
        return newDups;
     }
 
+    static byte[] buffer = new byte[1024];
+
     static private long getChecksum (File f) throws IOException
     {
        //java.util.zip.CRC32
@@ -108,7 +110,8 @@
        CheckedInputStream in = new CheckedInputStream (
            new BufferedInputStream (new FileInputStream (f)),
            checksum);
-       while (in.read () != -1) ;
+       while (in.read (buffer, 0, 1024) != -1) ;
+
        return checksum.getValue ();
     }

Now I get:

$ /usr/bin/time ./dupf /usr/lib/* > /dev/null
2.11user 0.43system 0:02.54elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1326major+624minor)pagefaults 0swaps

with GCJ, and:

$ /usr/bin/time java DupFinder /usr/lib/* > /dev/null
4.46user 0.53system 0:05.20elapsed 95%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (1598major+925minor)pagefaults 0swaps

With HotSpot client. I'd also reccomend using ArrayList instead of 
Vector (its faster due to no synchronization), and likewise HashMap 
instead of Hashtable.

> I've heard horror stories about upgrading the gcc, glibc or related 
> components from Red Hat. May I get just gcj sources and use them 
> without fear or should I look for newer rpms from redhat? 


Its quite safe to build GCC from source and install it to a different 
place (configure --prefix=/my/path ...) thus not overwriting the default 
compiler.

regards

Bryce.