problems with Deuce STM
Guy Korland
(BTW, the "sup...@deucestm.org" e-mail address does not work.)
I'm trying to make the STMBench7 benchmark run with Deuce STM, and I
have (so far) two main issues. It would be great if you knew how to
solve those.
1. java.* classes
STMBench7 uses several classes from java.*, e.g., java.util.Random and
various collections. Deuce STM does not seem to instrument those (I
get
exceptions about missing methods/fields). Is there any workaround that
does not require re-implementing all those standard classes?
2. Enumerations
Doing something like the following:
switch(e) {
A: ...
B: ...
C: ...
}
does not work if "e" is an Enum. The suggested workaraound (from the
mailing list):
switch(((Enum)e).ordinal()) {
0: ...
1: ...
2: ...
}
works but is ugly, because I have to replace labels by the
corresponding
numbers. Is there another way? I'd really like to avoid changing all
the
"switch" statements in the benchmark.
Best,
Michal
Guy Korland
The upcoming version 1.1 should solve it, providing offline
instrumentation to the java.* classes.
The current solution can be adding the classes to the exclude, e.g.
-Dorg.deuce.exclude="java.*,sun.*,org.eclipse.*,org.junit.*,junit.*"
see: http://www.deucestm.org/documentation/writing-your-own-stm-algorithm
>2. Enumerations
I'm still working on this bug and I hope I'll fix it soon, but I'm not
sure it will be soon, sorry...
Guy
On Oct 15, 8:58 am, Guy Korland <gkorl...@gmail.com> wrote:
> Hi Guy,
>
> (BTW, the "supp...@deucestm.org" e-mail address does not work.)
Guy Korland
java.util.TreeMap in the exclude list, then I won't be able to safely
access it from concurrent transactions, will I?
Michal
Guy Korland
The current version in the SVN already support full offline
instrumentation.
Meaning if instrument the rt.jar and your jars it should work fine.
Guy
Guy Korland
for the off-line instrumentation of rt.jar?
Guy Korland
for the off-line instrumentation of rt.jar?
Guy Korland
Guy
Guy Korland
Guy
Guy Korland
Guy
Michal Kapalka
java -Dorg.deuce.exclude="" -javaagent:deuceAgent.jar stmbench7.Benchmark
I get the exception:
java.lang.NoSuchFieldError: err__ADDRESS__
when the benchmark executes System.err.println. Am I missing some
important command-line option here?
Michal
Guy Korland
it offline instrument the rt.jar and your jars.
See: http://www.deucestm.org/documentation/getting-started
Guy
--
Guy Korland
Michal Kapalka
before on the "getting started" page...)
I still have a few problems, though. I pushed calls to System.err out of
transactions (where is their place, in fact), and instrumented off-line
both rt.jar and the jar of the benchmark. But I get errors, depending on
the value of org.deuce.exclude.
1. org.deuce.exclude=""
=============== DEBUG MESSAGE: Bad state ================
Error occurred during initialization of VM
java.lang.ExceptionInInitializerError
Caused by: java.lang.NullPointerException
at java.lang.String.charAt(String.java:689)
at java.security.BasicPermission.init(BasicPermission.java:93)
at java.security.BasicPermission.<init>(BasicPermission.java:129)
at java.lang.RuntimePermission.<init>(RuntimePermission.java:335)
at java.lang.Thread.<clinit>(Thread.java:1493)
2. org.deuce.exclude="java.lang.*"
Error occurred during initialization of VM
java.lang.NullPointerException
at java.util.Hashtable.put(Hashtable.java:394)
at java.lang.System.initProperties(Native Method)
at java.lang.System.initializeSystemClass(System.java:1072)
3. default org.deuce.exclude
Obviously, this leads to a NoSuchMethodError exception within the first
transaction.
Any hints would be welcome.
Best,
Michal
Guy Korland
fix it as soon as possible.
Guy
--
Guy Korland
Guy Korland
It seems like this version fixes these issues.
Guy
--
Guy Korland
Michal Kapalka
problems:
1. I have to put "sun.*" in the exclude list; otherwise, the benchmark
does not even start (no messages; it just hangs when starting the VM).
2. With "sun.*" in the exclude list I get the following exception:
Exception in thread "Thread-1" java.lang.IncompatibleClassChangeError:
Expecting non-static method
java.lang.Thread.currentThread()Ljava/lang/Thread;
at java.lang.Thread.currentThread(Thread.java)
at java.lang.ThreadLocal.get(ThreadLocal.java:124)
(...)
Any idea why this happens?
Note that I cannot put "java.lang.*" in the exclude list, because I
need, e.g., the Iterable interface to get instrumented -- otherwise I
get NoSuchMethodError exceptions in the benchmark.
Best,
Michal
Guy Korland
Guy Korland
Michal Kapalka
Exception in thread "Thread-1" java.lang.NoSuchFieldError: values__ADDRESS__
at java.util.TreeMap.values(TreeMap.java:817)
(...)
Maybe I am doing something wrong. Here are the commands that I use to
instrument and run the benchmark:
java -cp deuceAgent-1.1.0.jar \
-Dorg.deuce.exclude='sun.*,java.lang.Thread' \
org.deuce.transform.asm.Agent \
rt.jar deuce_rt.jar
java -cp deuceAgent-1.1.0.jar \
-Dorg.deuce.exclude='sun.*,java.lang.Thread' \
org.deuce.transform.asm.Agent \
stmbench7.jar deuce_bench.jar
java -Dorg.deuce.exclude='sun.*,java.lang.Thread' \
-cp deuceAgent-1.1.0.jar:deuce_bench.jar \
-Xbootclasspath/p:deuce_rt.jar:deuceAgent-1.1.0.jar \
stmbench7.Benchmark \
-g stm \
-s stmbench7.impl.deucestm.DeuceSTMInitializer
Guy Korland
BTW, a fix for the Enum.ordinal() was just committed to the trunk.
See: http://code.google.com/p/deuce/issues/detail?id=40
Guy
Guy Korland
Guy Korland
Which version of stmbench7 are you using?
Can you send me a link to the jar?
Guy
Guy Korland
Guy Korland
I think I found the problem I'm still verifing it.
Can you please try the new 1.2m2? http://www.deucestm.org/download/deuceAgent-1.2.0m1.jar
I ran it for now with:
-Dorg.deuce.exclude="java.lang.Enum,sun.*"
Thanks,
Guy
On Oct 24, 11:58 pm, Guy Korland <gkorl...@gmail.com> wrote:
> I'll try to reproduce it.
> Which version of stmbench7 are you using?
> Can you send me a link to the jar?
>
> Guy
>
Michal Kapalka
Thanks for the new version. Now the benchmark finishes; however, it
reports consistency (invariants) violations. More precisely, it seems
that at some point during the benchmark run the non-final String
attribute "type" gets cut in some objects. This attribute must be of the
form "type #(num)" where (num) is an integer number, but it becomes
"type #". The strange thing is that there is no operation that modifies
attribute "type" -- it could have been declared "final".
If you want to reproduce the result, run the benchmark and look for
exceptions above the "Detailed results" section of the benchmark output.
Or redirect stdout to /dev/null (all errors are sent to stderr).
Guy Korland
Guy Korland
Michal Kapalka
java '-Dorg.deuce.exclude=java.lang.Enum,sun.*' \
-cp deuceAgent-1.2.0m1.jar:deuce_bench.jar \
-Xbootclasspath/p:deuce_rt.jar:deuceAgent-1.2.0m1.jar \
-s stmbench7.impl.deucestm.DeuceSTMInitializer
Michal
>
> Guy
>
>
> I'll try to reproduce it.
> Which version of stmbench7 are you using?
> Can you send me a link to the jar?
>
> Guy
>
> On Thu, Oct 22, 2009 at 3:11 PM, Michal Kapalka
Guy Korland
--~--~---------~--~----~------------~-------~--~----~
You received this message because you are subscribed to the Google Groups "Deuce-STM developers" group.
To post to this group, send email to deuce-stm-...@googlegroups.com
To unsubscribe from this group, send email to deuce-stm-develo...@googlegroups.com
For more options, visit this group at http://groups.google.com/group/deuce-stm-developers?hl=en
-~----------~----~----~----~------~----~------~--~---
--
Guy Korland
Michal Kapalka
Yes, invariants check is executed without any transaction. But it's done
by a single thread, after the core benchmark code finishes, so I thought
it shouldn't be a problem. Initially, the invariants were checked inside
a transaction, but then I had problems with "System.err.println" giving
errors (see one of my first e-mails in this thread).
So here is what happens in the benchmark:
1. The data structure is set up and the invariants are checked: this is
done by a single thread, no transactions.
2. A number of threads execute transactions on the data structure.
3. Invariants are checked again by a single thread, no transactions.
The benchmark works well if I use locking instead of DeuceSTM in step 2
-- the same jar file, but different command-line parameters.
Is it a problem that the same data is accessed inside and outside of
transactions?
Thanks,
Michal
Guy Korland wrote:
> Hi,
>
> stmbench7.impl.DefaultOperationExecutor.execute(DefaultOperationExecutor.java:24)*
> * **at
> stmbench7.OperationExecutorFactory$1.run(OperationExecutorFactory.java:30)*
>
> Doesn't that meant that the Deuce's code isn't running but instead the
> Default version is running?
>
> Guy
>
>
> On Tue, Nov 3, 2009 at 12:25 PM, Michal Kapalka <michal....@epfl.ch
> <mailto:michal....@epfl.ch>> wrote:
>
>
> > Can you please send me the input arguments? and how to run the
> benchmark?
>
> Here is the command:
>
> java '-Dorg.deuce.exclude=java.lang.Enum,sun.*' \
> -cp deuceAgent-1.2.0m1.jar:deuce_bench.jar \
> -Xbootclasspath/p:deuce_rt.jar:deuceAgent-1.2.0m1.jar \
> stmbench7.Benchmark -g stm \
> -s stmbench7.impl.deucestm.DeuceSTMInitializer
>
> Michal
>
> >
> > Guy
> >
> > On Tue, Nov 3, 2009 at 12:07 PM, Michal Kapalka
> <michal....@epfl.ch <mailto:michal....@epfl.ch>
> >
> > I'll try to reproduce it.
> > Which version of stmbench7 are you using?
> > Can you send me a link to the jar?
> >
> > Guy
> >
> > On Thu, Oct 22, 2009 at 3:11 PM, Michal Kapalka
> > <michal.kapa...@epfl.ch
> <mailto:deuce-stm-developers%2Bunsu...@googlegroups.com>
> http://groups.google.com/group/deuce-stm-developers?hl=en
> -~----------~----~----~----~------~----~------~--~---
>
>
>
>
> --
> Guy Korland
>
> You received this message because you are subscribed to the Google
> Groups "Deuce-STM developers" group.