Oliver Deakin wrote:
Hi Gregory,
Regis made most of the performance improvements, so I will let him
answer on that front. From the point of view of bugs, here are some of
the rough areas we made fixes in:
- Stepping correctness - there were some bugs related to debug stepping,
particularly once we had JSR-45 support in and started debugging
non-Java stratum code. As an example, there were some scenarios where a
user would have to click step-over twice to move over a single line of
Java code. In non-Java code, there were cases where the line map would
not be interpreted correctly and we would end up "running away" as if we
had resumed the code rather than stepping.
- A number of crashes were uncovered, many from timing issues which were
exposed as the performance increased. These generally came from using
pointers that had already been freed, double freeing a pointer or
similar memory access/addressing issues.
- Some bugs related to JDWP returning incorrect values to queries. For
example, getting a count of all instances of a class when debugging in
Eclipse would often return an incorrect value, as (iirc) the agent was
holding references to objects that should have been released and made
available for garbage collection. The number returned would then appear
higher than it should because JDWP was still keeping instances alive
that were no longer referenced by the application.
- Issues around socket behaviour varying from platform to platform
created a number of bugs.
I hope this gives some idea of the kind of bugs we fixed - some related
to debug behaviour correctness, some related to porting to other
platforms and some were more general programming bugs (e.g. using freed
pointer).
Regards,
Oliver
Gregory Shimansky wrote:
This is very good news to me. I like to see this part of Harmony to
continue its development since JDWP agent wasn't touched for quite
some time since its contribution with just several bugs fixes by its
original developers and me.
Could you give the details about what actually increased the
performance in it and what kind of bugs were fixed?
On 29 April 2009 Oliver Deakin wrote:
Hi all,
A little while ago, IBM created a mirror of the Java 6 jdktools branch
so that we could start bringing Harmony JDWP into the IBM Java releases.
We began working in that mirror because it was convenient for our
internal builds, with a plan to reflect all changes we made back into
Harmony as we went. Unfortunately, as deadlines drew in and releases
came and went those plans did not come into fruition - until now :)
I'd like to announce the contribution of an enhanced Java 6 level JDWP
agent and socket transport layer. It has been attached to HARMONY-6187
[1] for everyone to try out - please take a look. It has the following
enhancements:
- A large number of bug fixes.
- Ported to a wide variety of platforms - Linux x86/x86_64, Linux PPC
32/64, zLinux 31/64, Windows x86/x86_64, AIX PPC 32/64 and zOS 31/64.
- As part of the porting process, all non-portable C++ usage has been
removed and rewritten, particularly use of C++ standard libraries.
- JSR 45 support implemented. The new JDWP agent allows debugging of
non-Java stratum code running on the VM, for example JSPs.
- Significant performance improvements.
All authors already have ACQs recorded. There may be some more work to
be done to get the new JDWP functioning 100% with Harmony, but I'd
rather do that work out in the open once it is committed. Please let me
know if there are any objections/comments to this contribution! I'll
give it a few days and then start a vote.
Regards,
Oliver
[1] https://issues.apache.org/jira/browse/HARMONY-6187
There are several performance improvements. Some could be treated as bugs, for
example, agent thread didn't manage JNI local reference (slow down jvmti call
which need to allocate memeory on java stack frame), thread wasn't removed from
thread list (slow down travelling list).
The most of improvements focus on memory manage and I/O:
- Use one worker thread to handle asynchronized command to avoid to start agent
thread every time
- Reduce times to read/write socket
- Use pre-allocated buffer for reading/writing data from/to socket to avoid to
allocate memory every time
- Optimize memory manage of PacketParser
--
Best Regards,
Regis.
