RFR for JDK-7190106 RMI benchmark fails intermittently because of use of fixed port
Stuart Marks
stuart.marks at oracle.com
Tue Nov 12 15:30:33 UTC 2013
Hi Tristan,
(Sorry, for the delay; I'm traveling)
I think the current package (bench.rmi and bench.serial) organization makes good
sense, and I don't want to change it. The need to shuffle things around to
different packages indicates a factoring error. In this case the culprit is
TestLibrary, which is in the unnamed package.
The constraints are:
(1) bench.rmi.* benchmarks depend on Main
(2) Main depends on TestLibrary
From (1), Main wants to be in bench.rmi. But (2) wants Main to be in the
unnamed package, since it can't get at TestLibrary otherwise.
Eventually we want to move TestLibrary into a named package. (This is
JDK-8003358). But this involves updating almost all of the RMI tests, so we
probably don't want to do that as part of this test fix.
The alternative is to put in a short-term fix (i.e., a hack :-)) to deal with
one of the dependencies. Since TestLibrary is causing the problem, which will
eventually be fixed, it makes sense to quick-fix the way Main depends on
TestLibrary. This would be cleaned up when TestLibrary gets moved into a named
package.
A class in a named package cannot import classes from the unnamed package, but
it can load them. So, one could leave the benchmarks and Main in bench.rmi, and
use a few lines of reflection code to load TestLibrary and to call
getUnusedRandomPort() on it. There's only one call to
TestLibrary.getUnusedRandomPort() from Main, so a few lines of reflective code
could be added right there.
So, that's my recommendation. If you do this you can leave all of the benchmarks
and the Main class in bench.rmi and avoid all the file moves.
Thanks,
s'marks
On 11/10/13 4:19 AM, Tristan Yan wrote:
> Hi Stuart
> I tried your suggestion but unfortunately all the benchmarks have dependencies
> to Main class because they need get stub from server. I suggest we move the
> benchmark tests to unnamed package unless we do want to put TestLibrary into a
> named package right now.
> Please let me know if you have objection on this.
> Thank you
> Tristan
>
> On 11/09/2013 02:28 AM, Stuart Marks wrote:
>> Hi Tristan,
>>
>> Yes, it's kind of a problem that the RMI TestLibrary is in the unnamed
>> package. Classes in a named package cannot import classes from the unnamed
>> package. We've run into problems with this before. Eventually, we should move
>> TestLibrary a named package.
>>
>> I think it's possible to work around this without too much difficulty. Note
>> that classes in the unnamed package can import classes from named packages.
>> So, perhaps you can put the RmiBench main class in the unnamed package so it
>> has access to TestLibrary. Then have the benchmarks themselves in the
>> bench.rmi package. The config file already references the benchmarks by fully
>> qualified class name (e.g., "bench.rmi.NullCalls") so with a bit of tinkering
>> you ought to be able to get this to work.
>>
>> s'marks
>>
>> On 11/8/13 3:00 AM, Tristan Yan wrote:
>>> Thank you, Stuart
>>> There is one review point I want to ask you opinion. Which is the reason that
>>> I moved from test/java/rmi/reliability/benchmark/bench/rmi to
>>> test/java/rmi/reliability/benchmark is Main.java need access class
>>> TestLibrary for supporting random port. TestLibrary is a unpackage class, I
>>> couldn't find a way to let a class which has Package to access the class
>>> without package. Do you have suggestion on that?
>>> Thank you so much.
>>> Tristan
>>>
>>> On 11/06/2013 09:50 AM, Stuart Marks wrote:
>>>>
>>>>
>>>> On 11/1/13 9:18 AM, Tristan Yan wrote:
>>>>> Hi Everyone
>>>>> http://cr.openjdk.java.net/~pzhang/Tristan/7190106/webrev/
>>>>>
>>>>> Description:
>>>>> 1. Convert shell script test to Java program test.
>>>>> 2. Using random server port by reusing Darryl Mocek's work to replace fixed
>>>>> server port.
>>>>> 3. Using java Process class to start client process.
>>>>> 4. Also convert other shell script test runSerialBench.sh to java program
>>>>> test also
>>>>
>>>> Hi Tristan,
>>>>
>>>> Several comments on this webrev.
>>>>
>>>>
>>>> ** The original arrangement within the test/java/rmi/reliability/benchmark
>>>> directory had the main benchmark files (scripts) at the top, some benchmark
>>>> framework files in the "bench" subdirectory, and the actual RMI and
>>>> serialization benchmarks in bench/rmi and bench/serial subdirectories.
>>>>
>>>> The webrev moves all the RMI benchmarks to the top benchmark directory but
>>>> leaves the serial benchmarks in bench/serial. The RMI benchmarks are now all
>>>> cluttering the top directory, but the main serial benchmark test is now
>>>> buried in the bench/serial directory. The nice organization that was there
>>>> before is now spoiled. Is this rearrangement necessary in order to convert
>>>> the scripts to Java? I would prefer the original arrangement be left in place.
>>>>
>>>>
>>>> ** The RMI benchmark Main.java file has a @run tag of the form,
>>>>
>>>> @run main/othervm/policy=policy.all/timeout=1800 -server Main -c config
>>>>
>>>> There is a subtle but serious problem here: the -server option is passed to
>>>> the >>JVM<< and not as an argument to the main() method. The main() method
>>>> gets neither a -server nor a -client argument, so its default "run mode" as
>>>> defined by the benchmark itself is SAMEVM. This runs the client and server
>>>> in the same JVM, which is different from the shell script, which ran
>>>> separate client and server JVMs.
>>>>
>>>>
>>>> ** The logic to process the -server argument still expects to take a port,
>>>> even though the port is assigned automatically. So the obvious fix to the
>>>> above,
>>>>
>>>> @run main/othervm/policy=policy.all/timeout=1800 Main -server -c config
>>>>
>>>> doesn't work, since a port is missing. The logic to increment the argument
>>>> index to collect the port argument should be removed. Also, the -server line
>>>> should be restored to the usage message, but without the port argument.
>>>>
>>>>
>>>> ** After this is done, the client's command line is constructed improperly.
>>>> The command line ends up looking something like this:
>>>>
>>>> java client -cp <classpath> Main client localhost:58583 -c config
>>>>
>>>> The word "client" appears twice, but what's really required is "-client" to
>>>> appear as an argument after Main.
>>>>
>>>>
>>>> ** The client is run using ProcessBuilder, which by default sends stdout and
>>>> stderr to pipes to be read by the parent. But the parent never reads them,
>>>> thus any messages from the client are never seen. The client is the one that
>>>> emits the benchmark report, so its output needs to be seen. It might be
>>>> sufficient to have the client inherit the parent's stdout and stderr. This
>>>> might intermix the client's and server's output, but it's better than nothing.
>>>>
>>>>
>>>> ** The config file is checked with the following code:
>>>>
>>>> try {
>>>> confFile = args[i];
>>>> confstr = new FileInputStream(System.getProperty("test.src")
>>>> + System.getProperty("file.separator") + confFile);
>>>> } catch (IOException e) {
>>>> die("Error: unable to open \"" + args[i] + "\"");
>>>> }
>>>>
>>>> This is potentially misleading, as the message doesn't print the actual
>>>> filename that was attempted to be opened.
>>>>
>>>> This is important, as the test.src property doesn't exist in the client JVM.
>>>>
>>>> Note that the original shell script passed full pathnames for the config
>>>> file to both the client and the server. The new @run tag merely says "-c
>>>> config" which redefines the config filename to be relative to the test.src
>>>> directory. You could pass -Dtest.src=... to the client, but it seems like
>>>> there should be something better than can be done.
>>>>
>>>>
>>>> ** The client needs to have its security policy set up. This is missing from
>>>> the construction of the client's command line.
>>>>
>>>>
>>>> ** ProcessBuilder takes a List<String> for its command; there is no need to
>>>> turn the list into an array.
>>>>
>>>>
>>>> ** In the benchmark main methods, code of the form,
>>>>
>>>> while (true) {
>>>> runBenchmarks();
>>>> if (exitRequested) {
>>>> System.exit();
>>>> }
>>>> }
>>>>
>>>> was replaced with
>>>>
>>>> while (!exitRequested) {
>>>> runBenchmarks();
>>>> }
>>>>
>>>> This is a subtle logic change, in that the former code always executed the
>>>> loop at least once. It seems unlikely, but it's possible that a timer could
>>>> set exitRequested before loop entry, resulting in the benchmark running zero
>>>> times. I guess, if you really want to clean this up (we do need to avoid
>>>> System.exit in jtreg tests), use a do-while loop instead.
>>>>
>>>>
>>>> ** Don't forget to remove the 7190106/runRmiBench.sh entry from
>>>> ProblemList.txt.
>>>>
>>>>
>>>> ** Remove the references to RMISecurityManager and just use SecurityManager.
>>>> This is just general cleanup. (I deprecated RMISecurityManager last week.) :-)
>>>>
>>>>
>>>> It would be good if you could fix up these issues and post another webrev.
>>>>
>>>> Thanks.
>>>>
>>>> s'marks
>>>>
>>>>
>>>>
>>>>
>>>>> Thank you
>>>>> Tristan
>>>>>
>>>>> On 01/11/2013 23:58, Stuart Marks wrote:
>>>>>> On 10/31/13 10:22 PM, Tristan Yan wrote:
>>>>>>> I am working on bug https://bugs.openjdk.java.net/browse/JDK-7190106. Based
>>>>>>> on my research, it looks like the issue of fixed port was already addressed
>>>>>>> by Stuart Marks in other RMI tests which are Java based. I would like to
>>>>>>> reuse his solution, however it does not work for shell based tests.
>>>>>>
>>>>>> (Darryl Mocek did the unique port work for the RMI tests.)
>>>>>>
>>>>>> Was the patch attached to your message? If so, it didn't get through. Most
>>>>>> OpenJDK mailing lists strip off attachments before forwarding the message to
>>>>>> the recipients.
>>>>>>
>>>>>>> 2. My recommendation would be to convert this shell script test into Java
>>>>>>> based test and re-use the dynamic port allocation solution by Stuart
>>>>>>> Marks to
>>>>>>> address the issue
>>>>>>>
>>>>>>> 3. Also this test was written with server/client mode in shell script. In
>>>>>>> the
>>>>>>> past there have been sync issues between server/client which caused the test
>>>>>>> to fail. If we convert the shell script into Java based test, it would avoid
>>>>>>> using "sleep 10" mechanism to allow for server and client to start up and
>>>>>>> also give us better control in synchronizing server and client.
>>>>>>
>>>>>> (Background for interested readers.) In general, yes, it's quite difficult to
>>>>>> make reliable shell tests, especially for multi-process tests like this one.
>>>>>> There is the unique port issue, and there is also the issue of how long for
>>>>>> the client to wait until the server is ready. Error handling is also a
>>>>>> problem, for example, if one of the JVMs gets an unexpected exception, it's
>>>>>> easy for shell tests to mishandle this case. They might hang or erroneously
>>>>>> report success.
>>>>>>
>>>>>> --
>>>>>>
>>>>>> If this is a rewrite, it's probably fairly large, so you need to upload it
>>>>>> somewhere (e.g., cr.openjdk.java.net) and then post a link to it.
>>>>>>
>>>>>> Thanks.
>>>>>>
>>>>>> s'marks
>>>>>
>>>
>>
>
More information about the core-libs-dev
mailing list