C++ vs Java "new" (no flame war please!)

[mlw]

Do not take anything about this, it is not a flame or troll, while I'm not
new to Java I favor C++. However, I may need to use it in a contract
position, and am concerned that the restrictions it places on application
code.

Take, for instance, this C++ construct:

class foo
{
char *m_name;
....
void * operator new(size_t size, char *string);
}
void *foo::new(size_ t size, char *string)
{
size_t cbstr = strlen(string)+ 1;
size_t cb = cbstr + size;
foo * t = (foo *) malloc(cb);
char * name = (char *) &t[1];
strcpy(name, string);
t->m_name = name;
}

The above example is a methodology that can be used to reduce the CPU and
memory overhead of malloc. You my argue that this is not a valid concern,
but if you have 10 or 100 million objects, the malloc block overhead,
alone, make this worth while. Hint: This is actually a simplification, some
times malloc is not used at all, and a big array is pre-alocated and work
through it with each new.
Is there a way to create 10 to 100 million objects in Java with a reasonable
system configuration?

[Lew]

mlw wrote:

(snipped description of custom 'new' operator)

Is there a way to create 10 to 100 million objects in Java with a reasonable
system configuration?

Sure, given the same considerations of available heap that you would have in
the C++ world.

Let's assume you want to create N objects where N is the largest number of
such objects that would fit in available memory.

List <Foostuff = new ArrayList <Foo(N);
for ( int x = 0; x < N; ++x )
{
stuff.add( new Foo( getAString() ) );
}
doSomethingWith ( stuff );

If you don't need them all in memory at once, it's even easier:

for ( int x = 0; x < N; ++x )
{
Foo foo = new Foo( getAString() );
doSomethingWith ( foo );
}

I posit the 'getAString()' method as where you'd obtain the equivalent of the
'char * string' in the C++ example. I assumed you'd use a different string for
each instance of Foo.

-- Lew

[mlw]

Lew wrote:

mlw wrote:

(snipped description of custom 'new' operator)

>Is there a way to create 10 to 100 million objects in Java with a
reasonable system configuration?

Sure, given the same considerations of available heap that you would have
in the C++ world.

Give or take, I guess.

>
Let's assume you want to create N objects where N is the largest number of
such objects that would fit in available memory.

List <Foostuff = new ArrayList <Foo(N);

That's one memory alloc, sure.

for ( int x = 0; x < N; ++x )
{
stuff.add( new Foo( getAString() ) );
}

The above code is exactly my problem with Java. In C++ I can overload new
and put all the objects anywhere I want, even in to one contiguous memory
block calling malloc merely once and combining the object and the string in
one allocation, for instance:

(Remember this is a simplification for example purposes, but the technique
is the important thing.)

unsigned char * myshared_block = shared_alloc(MA X_SIZE)
size_t curr_offset=0;

void *foo::operator new(size_t size, char * str)
{
size_t cbstr = strlen(str)+1;
size_t cb = size + cbstr;

foo * fooT = (foo *) &myshared_bl ock[curr_offset];
curr_offset += cb;
char *pstr = (char *)&foo[1];
strcpy(pstr, str);
fooT->str = pstr;
return (void *) fooT;
}

In the above code, I can pre-allocate a single memory block and pull objects
out of it until it is empty.

Every memory allocation has overhead, in GCC malloc, it is probably 4 bytes,
8 bytes on 64 bit systems. So, if you have fairly small objects, and lots
of them, a good chunk of memory will be eaten up with malloc overhead. If
you have a small object with a string, you will have two memory
allocations!

Obviously this is a rare problem, but it is a problem none the less.

doSomethingWith ( stuff );

>
If you don't need them all in memory at once, it's even easier:

for ( int x = 0; x < N; ++x )
{
Foo foo = new Foo( getAString() );
doSomethingWith ( foo );
}

I posit the 'getAString()' method as where you'd obtain the equivalent of
the 'char * string' in the C++ example. I assumed you'd use a different
string for each instance of Foo.

[Lew]

Lew wrote:

>Let's assume you want to create N objects where N is the largest number of
such objects that would fit in available memory.

List <Foostuff = new ArrayList <Foo(N);

mlw wrote:

That's one memory alloc, sure.

Lew wrote:

> for ( int x = 0; x < N; ++x )
{
stuff.add( new Foo( getAString() ) );
}

mlw wrote:

The above code is exactly my problem with Java. In C++ I can overload new
and put all the objects anywhere I want, even in to one contiguous memory
block calling malloc merely once and combining the object and the string in
one allocation,

You still need to calculate offsets into that block to fix the start of each
individual object. In fact, the code in your custom allocator uses more
memory and much more time than would the JVM for the equivalent Java class.

Every memory allocation has overhead, in GCC malloc, it is probably 4 bytes,
8 bytes on 64 bit systems. So, if you have fairly small objects, and lots
of them, a good chunk of memory will be eaten up with malloc overhead. If
you have a small object with a string, you will have two memory
allocations!

Your C++ class did a copy of the string. A Java class likely would not, since
Strings are immutable, so it would only do one allocation for the Foo object
and none for the String. Even if one did copy the String, the time overhead
of the Java allocation and copy would be much less than for the C++ code you
showed. For one thing, the Java code would only loop through the String once,
not twice as in your C++ code; it would have no need to calculate "strlen()".
The memory overhead would be no different since a copy is a copy is a copy.

But as I said, the Java code would not copy the String, so the point is moot.
One allocation and less memory overhead in the Java version.

Obviously this is a rare problem, but it is a problem none the less.

I don't see what the problem with Java is. What is the bad effect that
concerns you with Java?

Is it memory overhead? Objects in Java take up only as much space as they take.

Is it time overhead? Java object allocations run on the order of 10-20
machine cycles. Initialization of the object takes some time, perhaps, but
that would be true with your custom allocator as well.

Your description seems to delineate a problem with C++ that your custom
allocator handles, but I see nothing of this relevant to Java.

-- Lew

[Mark Rafn]

mlw <ml*@nospamnowa y.zzwrote:

>Take, for instance, this C++ construct:
void *foo::new(size_ t size, char *string)
{
size_t cbstr = strlen(string)+ 1;
size_t cb = cbstr + size;
foo * t = (foo *) malloc(cb);
char * name = (char *) &t[1];
strcpy(name, string);
t->m_name = name;
}

>The above example is a methodology that can be used to reduce the CPU and
memory overhead of malloc.

Trying not to flame C++ here, but I'm glad not to see very much of this kind
of code in Java.

>You my argue that this is not a valid concern

In some apps, it is. If you're extremely sensitive to exact memory
allocation, or need hardware access that Java doesn't give you (say, to SysV
shared memory or something), C++ is a good choice. I'd generally recommend to
do the specific sensitive bit in C++ and the rest in Java, but it'll depend
entirely on what the app actually does.

>Is there a way to create 10 to 100 million objects in Java with a reasonable
system configuration?

Any VM since 1.4 on modern hardware should not have a problem with this,
unless you're pretty time-sensitive.
--
Mark Rafn da***@dagon.net <http://www.dagon.net/>

[mlw]

Lew wrote:

Lew wrote:

>>Let's assume you want to create N objects where N is the largest number
of such objects that would fit in available memory.

List <Foostuff = new ArrayList <Foo(N);

mlw wrote:

>That's one memory alloc, sure.

Lew wrote:

>> for ( int x = 0; x < N; ++x )
{
stuff.add( new Foo( getAString() ) );
}

mlw wrote:

>The above code is exactly my problem with Java. In C++ I can overload new
and put all the objects anywhere I want, even in to one contiguous memory
block calling malloc merely once and combining the object and the string
in one allocation,

You still need to calculate offsets into that block to fix the start of
each
individual object. In fact, the code in your custom allocator uses more
memory and much more time than would the JVM for the equivalent Java
class.

That is simply not true on either account, I don't need to "fix" the start
of anything. There is no memory overhead for each class. The "allocation "
overhead time is nothing more than a simple integer addition.

>

>Every memory allocation has overhead, in GCC malloc, it is probably 4
bytes, 8 bytes on 64 bit systems. So, if you have fairly small objects,
and lots of them, a good chunk of memory will be eaten up with malloc
overhead. If you have a small object with a string, you will have two
memory allocations!

Your C++ class did a copy of the string.

Yes.

A Java class likely would not,
since Strings are immutable, so it would only do one allocation for the
Foo object
and none for the String.

Well, not seen in the example would be that the string is read from a file
into a local buffer which gets reused.

Even if one did copy the String, the time
overhead of the Java allocation and copy would be much less than for the
C++ code you
showed.

How is this possible?

char buffer[MAX_SIZE]

while(!feof(f))
{
if(fgets(buffer , sizeof(buffer), f))
{
new(buffer) foo();
}
}

Where is the allocation overhead that you are referring too?

For one thing, the Java code would only loop through the String
once, not twice as in your C++ code; it would have no need to calculate
"strlen()".
The memory overhead would be no different since a copy is a copy is a
copy.

Like I said, I was showing a technique that was simplified for example,
don't nit-pick the example because it is obviously much less sophisticated
then the actual application that uses the technique. In the real code, the
new operator is much more complex and reads the data itself.

>
But as I said, the Java code would not copy the String, so the point is
moot. One allocation and less memory overhead in the Java version.

Java would allocate a new string for each string. My code does not issue any
memory allocation for the string, it comes from a fixed length block and it
is combined with the object proper.

>

>Obviously this is a rare problem, but it is a problem none the less.

I don't see what the problem with Java is. What is the bad effect that
concerns you with Java?

Is it memory overhead? Objects in Java take up only as much space as they
take.

Try this:

allocate 20,000,000 objects with each object containing at least one string.
The memory footprint of the application has AT LEAST 160 megabytes of
overhead that can be virtually eliminated by pre-allocating 4 megabyte
blocks and sub-allocating out of that.

>
Is it time overhead? Java object allocations run on the order of 10-20
machine cycles. Initialization of the object takes some time, perhaps,
but that would be true with your custom allocator as well.

Java allocations take 10-20 machine cycles? Not on your life. In JIT
compiled code in which objects lose scope on exit of the function, yes,
that may be the overall time (it could even be much less), but objects that
live beyond a function scope have the overhead of a memory allocation.

>
Your description seems to delineate a problem with C++ that your custom
allocator handles, but I see nothing of this relevant to Java.

Java:

class foo
{
String m_test;

foo(String value)
{
m_test = value;
}
void print()
{
System.out.prin tln(m_test+"\n" );
}
};
public class test
{
public static void main(java.lang. String[] args)
{
int i=0;
try
{
foo arr[] = new foo[2000000];
for(i=0; i < 2000000; i++)
arr[i] = new foo(Integer.toH exString(i));
}
catch(OutOfMemo ryError e)
{
System.out.prin tln(e.getMessag e());
System.out.prin tln("Total object:" + i);
}
}
};
C++
#include <stdlib.h>
#include <unistd.h>
#include <assert.h>
#include <string.h>
#include <stdio.h>

class foo
{
char *m_test;

public:
foo();
void * operator new(size_t size, void *p, size_t cb);
};

#define BLKSIZE 1024*1024

unsigned char *block = NULL;
size_t blk_size=0;
size_t blk_offset=0;
foo::foo()
{
}
void *foo::operator new(size_t size, void *p, size_t cb)
{
size_t totalcb = size+cb;

if(!block || totalcb (blk_size - blk_offset))
{
block = (unsigned char *) malloc(BLKSIZE) ;
blk_size = BLKSIZE;
blk_offset = 0;
}
assert(block);

foo *fooT = (foo *) &block[blk_offset];
blk_offset += size;
fooT->m_test = (char *) &block[blk_offset];
blk_offset += cb;
memcpy(fooT->m_test, p, cb);
return (void *) fooT;
}

int main()
{
int i=0;
foo ** ar = (foo **) malloc(sizeof(f oo*) * 2000000);

for(i=0; i < 2000000; i++)
{
char buffer[64];
size_t cb = snprintf(buffer ,sizeof(buffer) , "%X", i)+1;
ar[i] = new((void *)buffer,cb) foo();
}
printf("%d\n", i);
}
The results:
test@localhost: ~/scat$ time java test
Java heap space
Total object:914828

real 0m23.519s
user 0m21.009s
sys 0m2.004s
test@localhost: ~/scat$ time ./test
2000000

real 0m0.838s
user 0m0.724s
sys 0m0.064s

[Lew]

mlw wrote:

That is simply not true on either account, I don't need to "fix" the start
of anything. There is no memory overhead for each class. The "allocation "
overhead time is nothing more than a simple integer addition.

It is that addition to which I refer.

Lew wrote:

>Even if one did copy the String, the time
overhead of the Java allocation and copy would be much less than for the
C++ code you
showed.

mlw wrote:

How is this possible?

I was referring to your code:

void *foo::new(size_ t size, char *string)
{
size_t cbstr = strlen(string)+ 1;
size_t cb = cbstr + size;
foo * t = (foo *) malloc(cb);
char * name = (char *) &t[1];
strcpy(name, string);
t->m_name = name;
}

strlen() has overhead and strcpy() has overhead.

Well, not seen in the example would be that the string is read from a file
into a local buffer which gets reused.

In Java, it would be read into a new buffer each time, the allocation overhead
of which is negligible, and the copy from the file into the buffer would be
the only copy. There would not be the copy represented in your code by the
strcpy() call. The small overhead of the new buffer allocation is well offset
by the savings in the string copy time.

char buffer[MAX_SIZE]

while(!feof(f))
{
if(fgets(buffer , sizeof(buffer), f))
{
new(buffer) foo();
}
}

Where is the allocation overhead that you are referring too?

in the new operator that you wrote, particularly in the strlen() and strcpy()
calls, which add to the time and memory footprints.

Java would allocate a new string for each string. My code does not issue any
memory allocation for the string, it comes from a fixed length block and it
is combined with the object proper.

But the Java code would have one less copy of that string, and that would save
copy time. The fact that the memory is allocated at new time in Java is about
the same as the memory add at new time in your C++ example. A Java object
allocation is not much more than a memory limit add.

arr[i] = new foo(Integer.toH exString(i));

size_t cb = snprintf(buffer ,sizeof(buffer) , "%X", i)+1;

You might be comparing the speed of toHexString() to that of snprintf(), and
not allocation times.

You also are not comparing Java with custom allocators to Java without custom
allocators. I am not arguing that Java is faster than C++, only that custom
allocators in Java would not help the Java performance.

With Java you have the overhead of bytecode interpretation and multiple
threads running in the JVM at the same time, plus there is the startup time of
the JVM itself that you did not factor out. Consequently you have not
measured memory allocation time vs. memory allocation time and we can draw no
conclusions about the relative efficiency of the two schemes.

Put your timing loops inside the program, and while you're at it give the Java
Hotspot compiler a few hundreds of loops to settle in. Oh, and try java
-client vs. java -server. You've got to factor out the overhead of setup and
so on before timing comparisons make sense.

-- Lew

[mlw]

Lew wrote:

mlw wrote:

>That is simply not true on either account, I don't need to "fix" the
start of anything. There is no memory overhead for each class. The
"allocation " overhead time is nothing more than a simple integer
addition.

It is that addition to which I refer.

Lew wrote:

>>Even if one did copy the String, the time
overhead of the Java allocation and copy would be much less than for the
C++ code you
showed.

mlw wrote:

>How is this possible?

I was referring to your code:

>void *foo::new(size_ t size, char *string)
{
size_t cbstr = strlen(string)+ 1;
size_t cb = cbstr + size;
foo * t = (foo *) malloc(cb);
char * name = (char *) &t[1];
strcpy(name, string);
t->m_name = name;
}

strlen() has overhead and strcpy() has overhead.

Only CPU overhead, and like I said, this is a simplification.

>

>Well, not seen in the example would be that the string is read from a
file into a local buffer which gets reused.

In Java, it would be read into a new buffer each time, the allocation
overhead of which is negligible, and the copy from the file into the
buffer would be
the only copy. There would not be the copy represented in your code by
the
strcpy() call. The small overhead of the new buffer allocation is well
offset by the savings in the string copy time.

>char buffer[MAX_SIZE]

while(!feof(f) )
{
if(fgets(buffer , sizeof(buffer), f))
{
new(buffer) foo();
}
}

Where is the allocation overhead that you are referring too?

in the new operator that you wrote, particularly in the strlen() and
strcpy() calls, which add to the time and memory footprints.

strlen and strcpy never call malloc. The function strdup does call malloc.

>

>Java would allocate a new string for each string. My code does not issue
any memory allocation for the string, it comes from a fixed length block
and it is combined with the object proper.

But the Java code would have one less copy of that string, and that would
save
copy time. The fact that the memory is allocated at new time in Java is
about
the same as the memory add at new time in your C++ example. A Java object
allocation is not much more than a memory limit add.

> arr[i] = new foo(Integer.toH exString(i));

size_t cb = snprintf(buffer ,sizeof(buffer) , "%X", i)+1;

You might be comparing the speed of toHexString() to that of snprintf(),
and not allocation times.

I think you are very confused about memory allocation in Java, it is not
trivial when it exceeds certain limits or exists outside function scope.
Just because you have no control over how Java does this, does not mean it
is something you don't need to know about.

When I change the lines to :
arr[i] = new foo("Test:"+i);
and
size_t cb = snprintf(buffer ,sizeof(buffer) , "Test:%X", i)+1;

In the Java and C++ code, respectively, I get pretty much the same results:
test@localhost: ~/scat$ time ./test
2000000

real 0m1.020s
user 0m0.860s
sys 0m0.092s
test@localhost: ~/scat$ time java test
Java heap space
Total object:741838

real 0m19.178s
user 0m16.893s
sys 0m1.728s

[mlw]

Mark Rafn wrote:

mlw <ml*@nospamnowa y.zzwrote:

>>Take, for instance, this C++ construct:
void *foo::new(size_ t size, char *string)
{
size_t cbstr = strlen(string)+ 1;
size_t cb = cbstr + size;
foo * t = (foo *) malloc(cb);
char * name = (char *) &t[1];
strcpy(name, string);
t->m_name = name;
}

>>The above example is a methodology that can be used to reduce the CPU and
memory overhead of malloc.

Trying not to flame C++ here, but I'm glad not to see very much of this
kind of code in Java.

Its not pretty, not at all, but it is the sort of code that can make the
difference between running or not running when necessary or withing the
time requirements. Fortunately, you can write it, test it, make sure it
works, then hide it in a library where newbees can't screw around and break
it.

>

>>You my argue that this is not a valid concern

In some apps, it is. If you're extremely sensitive to exact memory
allocation, or need hardware access that Java doesn't give you (say, to
SysV
shared memory or something), C++ is a good choice. I'd generally
recommend to do the specific sensitive bit in C++ and the rest in Java,
but it'll depend entirely on what the app actually does.

That's sort of why I posted the thread. I may need to do some "interestin g"
things in Java and was kind of looking to see if anyone could come up with
a good trick or two.

>

>>Is there a way to create 10 to 100 million objects in Java with a
reasonable system configuration?

Any VM since 1.4 on modern hardware should not have a problem with this,
unless you're pretty time-sensitive.

To address the "time-sensitive" comment, many times I hear that the machines
are fast enough that you don't need to worry about performance, I have to
say this is the same argument for the 4 day work week. "Soon, we'll be
productive enough that we'll only have to work 4 days a week." That was the
dream and the myth. The problem with that line of thought is that as
capability is increased, so are expectations. I don't know about you, but
the 4 day work week hasn't come to my corner of the globe.

If you make a program that takes an hour to do something, and someone comes
along and writes one that takes 5 minutes, you lose, no matter what the
program is written in.

[Lew]

mlw wrote:

real 0m1.020s
user 0m0.860s
sys 0m0.092s
test@localhost: ~/scat$ time java test
Java heap space
Total object:741838

real 0m19.178s
user 0m16.893s
sys 0m1.728s

You're still not timing memory allocation but JVM startup and other factors.

-- Lew