Java slow on Xeon processors ?

[Michael Kreitmann]

Hello,

I've a strange performance problem with the follwoing (senseless)
code:

public class XeonTest
{

public static String XMLEncode(String aText)
{
int len = aText.length();
StringBuffer result = new StringBuffer(len);
char ch;
for (int i = 0; i < len; ++i)
{
ch = aText.charAt(i);
result.append(ch);
}
return result.toString();
}

public static void main(String[] args)
{
long start = System.currentTimeMillis();
String s = "The brown fox jumps over the bridge. The brown fox
jumps over the bridge";
String res = "";
for (int i = 0; i < 1000000; ++i)
{
res = XMLEncode(s);
}
System.out.println ("Runtime:" +
(System.currentTimeMillis()-start)+" ms" );
}
}

I'm using java 1.4.2 with the "-server" parameter (but this doesnt'
matters) and Windows 2000 as OS.
Time comparision:
PIII/1000: 5 seconds
2xXeon/1800: 14 seconds
PIV/3000/: 2 seconds

All other performance test on the Xeon machine (SiSoft Sandra) are
runnig good.
Can anybody please give me a hint, what's going wrong here ? All the
time is lost in the "result.append(ch)" line. But why is the Xeon so
slow ?
It would also be great, if anybody could let run the code on any other
Intel Xeon machine ;-)

Many thanks for your help!

Regards
Michael

 

[Roedy Green]

Can anybody please give me a hint, what's going wrong here ? All the
time is lost in the "result.append(ch)" line. But why is the Xeon so
slow ?

One possibility is the Xeon is not particularly good at character
addressing. It has to read, poke, write to store a byte. So do other
processors, but maybe it is more optimised for big chunks.

Try a similar benchmark that works with 16 32 and 64 bit quantities.

 

[Nigel Wade]

Hello,

I've a strange performance problem with the follwoing (senseless)
code:

public class XeonTest
{

public static String XMLEncode(String aText)
{
int len = aText.length();
StringBuffer result = new StringBuffer(len);
char ch;
for (int i = 0; i < len; ++i)
{
ch = aText.charAt(i);
result.append(ch);
}
return result.toString();
}

public static void main(String[] args)
{
long start = System.currentTimeMillis();
String s = "The brown fox jumps over the bridge. The brown fox
jumps over the bridge";
String res = "";
for (int i = 0; i < 1000000; ++i)
{
res = XMLEncode(s);
}
System.out.println ("Runtime:" +
(System.currentTimeMillis()-start)+" ms" );
}
}

I'm using java 1.4.2 with the "-server" parameter (but this doesnt'
matters) and Windows 2000 as OS.
Time comparision:
PIII/1000: 5 seconds
2xXeon/1800: 14 seconds
PIV/3000/: 2 seconds

All other performance test on the Xeon machine (SiSoft Sandra) are
runnig good.
Can anybody please give me a hint, what's going wrong here ? All the
time is lost in the "result.append(ch)" line. But why is the Xeon so
slow ?
It would also be great, if anybody could let run the code on any other
Intel Xeon machine ;-)

Many thanks for your help!

Regards
Michael

If I were to speculate, I'd say it was most likely due to the system being
dual-processor. It might be due to some internal synchronization within
StringBuffer which slows the code down considerably on multi-processor machines.

Just so you know it's not a Windows problem, I see exactly the same
difference between a 2.8MHz P4 and dual 2.4GHz Xeon running RedHat FC1.
For the mono processor the code runs in 2300-2350ms, on the dual Xeon it
runs in ~11000ms. I do notice that while the code is running on the dual
processor system that both CPUs are showing around 30-50% utiliztion
indicating an inablitity to get 100%.

It would be interesting to see how the code executes on a single processor Xeon.

 

[Dave]

I've a strange performance problem with the follwoing (senseless)

code:

I'm using java 1.4.2 with the "-server" parameter (but this doesnt'
matters) and Windows 2000 as OS.
Time comparision:
PIII/1000: 5 seconds
2xXeon/1800: 14 seconds
PIV/3000/: 2 seconds

All other performance test on the Xeon machine (SiSoft Sandra) are
runnig good.

Hello Michael,

When you say all other performance tests are running OK, do you mean all
other Java performance tests? Or non-Java tests?

Which JVM?
Is compiler optimization on?
Is JIT on?

If this is the only Java test, and you are not using the JIT option and
not compiling with optimization on, that might account for the
difference...

HTH

--Dave

 

[Michael Kreitmann]

Hello,

I've a strange performance problem with the follwoing (senseless)
code:

public class XeonTest
{

public static String XMLEncode(String aText)
{
int len = aText.length();
StringBuffer result = new StringBuffer(len);
char ch;
for (int i = 0; i < len; ++i)
{
ch = aText.charAt(i);
result.append(ch);
}
return result.toString();
}

public static void main(String[] args)
{
long start = System.currentTimeMillis();
String s = "The brown fox jumps over the bridge. The brown fox
jumps over the bridge";
String res = "";
for (int i = 0; i < 1000000; ++i)
{
res = XMLEncode(s);
}
System.out.println ("Runtime:" +
(System.currentTimeMillis()-start)+" ms" );
}
}

I'm using java 1.4.2 with the "-server" parameter (but this doesnt'
matters) and Windows 2000 as OS.
Time comparision:
PIII/1000: 5 seconds
2xXeon/1800: 14 seconds
PIV/3000/: 2 seconds

All other performance test on the Xeon machine (SiSoft Sandra) are
runnig good.
Can anybody please give me a hint, what's going wrong here ? All the
time is lost in the "result.append(ch)" line. But why is the Xeon so
slow ?
It would also be great, if anybody could let run the code on any other
Intel Xeon machine ;-)

Many thanks for your help!

Regards
Michael

If I were to speculate, I'd say it was most likely due to the system being
dual-processor. It might be due to some internal synchronization within
StringBuffer which slows the code down considerably on multi-processor machines.

Just so you know it's not a Windows problem, I see exactly the same
difference between a 2.8MHz P4 and dual 2.4GHz Xeon running RedHat FC1.
For the mono processor the code runs in 2300-2350ms, on the dual Xeon it
runs in ~11000ms. I do notice that while the code is running on the dual
processor system that both CPUs are showing around 30-50% utiliztion
indicating an inablitity to get 100%.

It would be interesting to see how the code executes on a single processor Xeon.

First of all: Thanks for your answers!
I've done another test on my P4 with activated Hyperthreading and ...
the performance is very bad (9s <> 2s without HT).
(The test on the 1xXeon/1800 will hopefully come tomorrow!)

So: Java (?) does have a problem running on dual-processor machines ?
Can this be right ? I can't believe, but what can I do ??

Regards
Michael

 

[Thomas Weidenfeller]

First of all: Thanks for your answers!
I've done another test on my P4 with activated Hyperthreading and ...
the performance is very bad (9s <> 2s without HT).
(The test on the 1xXeon/1800 will hopefully come tomorrow!)

If you have java 1.5 available, try to replace StringBuffer with
StringBuilder. The later is not synchronized, so should not run into the
same problem.

/Thomas

 

[Timo Kinnunen]

If you have java 1.5 available, try to replace StringBuffer with
StringBuilder. The later is not synchronized, so should not run into the
same problem.

Another thing to try on a pre 1.5 Java is to put the whole loop inside a
synchronized block, synchronized with the StringBuffer.

 

[Juha Laiho]

....
Just so you know it's not a Windows problem, I see exactly the same
difference between a 2.8MHz P4 and dual 2.4GHz Xeon running RedHat FC1.
For the mono processor the code runs in 2300-2350ms, on the dual Xeon it
runs in ~11000ms. I do notice that while the code is running on the dual
processor system that both CPUs are showing around 30-50% utiliztion
indicating an inablitity to get 100%.

Hmm.. As I see the code, it's a single-threaded app, so should ideally
take 100% of the processing power of one CPU. If it takes 50% of power
on 2 CPUs, this is somewhat equal, minus the frequent need to clear
CPU caches and synchronise CPU states, but that is an inefficiency of
the OS scheduler (not able to bind a CPU-looping process to a single CPU).

 

[Roedy Green]

So: Java (?) does have a problem running on dual-processor machines ?
Can this be right ? I can't believe, but what can I do ??

you are running a single thread, so the extra processor is not likely
the problem. If StringBuffer is thread safe, it could however be the
extra overhead of thread safety on the Xeon.

Replace the code with a dummy non-thread safe String buffer that just
does

void append ( char c )
{
array ( ++i ) = c;
}

and see if you still see the anomaly. If you do, the culprit is likely
byte addressing. If you don't, it is likely sync overhead. You can
then experiment further adding more code from the real StringBuffer.

StringBuffers that are not initialised to a decent starting size are
inefficient the way they gradually grow, leaving behind droppings of
discarded incarnations.

 

[Roedy Green]

Hmm.. As I see the code, it's a single-threaded app, so should ideally
take 100% of the processing power of one CPU. If it takes 50% of power
on 2 CPUs, this is somewhat equal, minus the frequent need to clear
CPU caches and synchronise CPU states, but that is an inefficiency of
the OS scheduler (not able to bind a CPU-looping process to a single CPU).

That's reasonable, so long at is it is giving decently long time
slices. Even on a single CPU, you would get that same overhead since
the OS still hands out short time slices to CPU bound tasks.


Binding to a single CPU in general would give worse performance.

 

[Michael Kreitmann]

Hello Michael,

When you say all other performance tests are running OK, do you mean all
other Java performance tests? Or non-Java tests?

Which JVM?
Is compiler optimization on?
Is JIT on?

If this is the only Java test, and you are not using the JIT option and
not compiling with optimization on, that might account for the
difference...

HTH

--Dave

Hello Dave,

I'm using Sun Java SDK 1.4.2 (1.4.2_04). The source is complied with
javac -O
The performance tests I mean are the tests included in SiSoft Sandra
2004. (Memory bandwith, CPU and so on).
JIT ? Hmm, how can I switch it on/off ? (Is this a very silly question
?)

Thanks for your advice!
Regards
Michael

 

[Marc Slemko]

First of all: Thanks for your answers!
I've done another test on my P4 with activated Hyperthreading and ...
the performance is very bad (9s <> 2s without HT).
(The test on the 1xXeon/1800 will hopefully come tomorrow!)

So: Java (?) does have a problem running on dual-processor machines ?
Can this be right ? I can't believe, but what can I do ??

Do you have to do anything?

Do you have reason to think this is a performance issue in your
code caused by this?

What your code is doing is essentially 35 million synchronized
method calls per second. Yes, synchronization is typically more
expensive on a multiproc system.

With some JVMs on some systems, adding a synchronized(result) block
around the whole result.append() loop could reduce this overhead;
with other JVMs it may hurt.

If you need this case to perform well, you need something like the
StringBuilder class in 1.5 that isn't synchronized. The typical
app, however, isn't going to be doing 35 million StringBuffer.append()
calls per second, so it becomes a very trivial difference in
performance.

I can reproduce the behaviour you desribe, if I change to a
StringBuilder then a dual 2.4 GHz Xeon is slightly faster than a
single 2.4 GHz P4, which is pretty much exactly what is expected.

There are all sorts of interesting differences... why the 1.5 beta JVM
is slower on a single proc system ... why the IBM 1.4 JDK is slower if
you add the synchronized block I mention ... but the bottom line
is that there are very few applications where it matters. For some
it does, which is the reason why the unsynchronized StringBuilder
was added to 1.5.

Performance is important, yes, but in context.