Performance/Meetings/2009 04
Performance/Meetings/2009/04/17
Meeting Minutes
IRC Meeting of Sun Microsystems (StarOffice) with RedFlag2000
Performance Project
Date: 2009/04/17
Time: 15:50– 16:27
Meeting No.:
Agenda:
(3:57:23 PM) peter13j: Good morning|afternoon everybody
(4:00:26 PM) liangjun: hello all:)
(4:00:48 PM) liheng: Good morning / afternoon
(4:00:49 PM) arwe: Hi All :-)
(4:00:56 PM) os_ooo: Hello!
(4:01:08 PM) liheng: Good morning / afternoon
(4:01:37 PM) xiuzhi: hi all
(4:01:42 PM) odf-mib: Hi all
(4:01:44 PM) yugq: hi all
(4:02:44 PM) cd_oo: hi all
(4:02:50 PM) erAck: good morning|afternoon
(4:03:02 PM) kuangliang: hi all
(4:03:26 PM) liheng: Now, we update overall status.
(4:06:32 PM) erAck: I'm about finishing Calc's area broadcasters per sheet separation. A constructed "very bad case" scenario shows improvement by 75% (read: 25% of previous time needed).
(4:06:38 PM) sb: my status: with ~every possible for improving Linux library loading done, now looking at ways to further decrease cold start time by reducing (data) file access; got side-tracked by looking at potential improvements for rtl::OUString (once again...)
(4:06:59 PM) xiuzhi: zhangyuwei and JackieSun can not join today's meeting. zhangyuwei is fixing thread lock issue, JackieSun will finish the coding of big images process
(4:07:37 PM) os_ooo: Bjoern Michaelsen improved the access to frames/graphics/OLE objects and shapes in xmloff. This makes saving of documents containing many of those objects significantly. In one of our test documents it makes about 20%. The cws is currently beeing rebased and will be brought to QA soon.
(4:08:43 PM) Matthias: Hi all, sorry for being late
(4:10:00 PM) yugq: yugq status: I finished a incremeantal save demo for odt file. The demo is based on standard C++ and libxml2, use quite similar data structure of inner OOo document structure. It's just a research on the possibility for incremental save of odt. Now, I focus on research the OOo writer document builder and find out a way to implement the increment save.
(4:12:09 PM) cd_oo: cd_oo status: I verified the findings from Yuan Cheng and can confirm that non-rebased libraries indeed boost cold start up performance on Windows. Unfortunately non-rebased libraries have other drawbacks and I am currently investigate how severe they are. At least non-rebased libraries don't share the code section between processes. Memory consumption is also higher on startup (normally...
(4:12:11 PM) cd_oo: ...most of the library must be loaded for the relocation process). I could verify that putting all libraries next to the application executable gives us a 10% cold start performance boost (OOo 3.0 has three folders for libraries).
(4:12:13 PM) cd_oo: The current state of my investigation can be found here:
(4:12:15 PM) cd_oo: http://wiki.services.openoffice.org/wiki/Performance/Reorder_Symbols_For_Libraries
(4:12:16 PM) cd_oo: http://wiki.services.openoffice.org/wiki/Performance/Library_and_directory_structure
(4:12:29 PM) arwe: 3.1 release struck me again, so i could not do too much abstract. One of the concrete cases which came up was the copy constructor and assigment operator of GraphicObject, which unnecessarily calculates an ObjectID using a Checksum. This is obviously not needed and brings speedups in many situations, including load/save.
(4:14:10 PM) liheng: I setuped glassfish and transformed a part of benchmark-system on server, and developed a commandline tools that has same function as Web Benchmark System, and update code to http://performance.openoffice.org/svn/performance/trunk/benchmark-system/src/tools
(4:14:50 PM) liangjun: Implement a series of simple applications, and get a series of test data .
(4:15:13 PM) liangjun: I will analyse the test results and put on wiki.
(4:16:51 PM) ***Matthias continues with research on buffered file I/O; currently reorganizing (refactoring) sal/osl/unx/file*.cxx for integration of buffering.
(4:18:23 PM) ***Matthias also had a (short) discussion with MAV about images.zip; independently tora-japan had similar ideas :-)
(4:22:27 PM) liheng: If no other topics, it seems like we can finish this meeting ahead of time. Any topic? :)
(4:23:11 PM) Matthias: nothing special from me; I'm coding ... :-)
(4:24:04 PM) liheng: Can we finished?
(4:24:35 PM) ***erAck says: let's ditch all API and use UTF-8 instead of UTF-16 ;-)
(4:24:42 PM) erAck: ok, see you next week.
(4:25:08 PM) Matthias: yes, have a nice weekend; see you next week; bye all
(4:25:25 PM) odf-mib: Good bye!
(4:25:33 PM) liheng: Okay,bye all!
(4:25:42 PM) arwe: Okay, see You :-)
(4:25:50 PM) liangjun: :) bye
(4:26:12 PM) yugq: bye all
(4:26:55 PM) cd_oo: Bye all
Performance/Meetings/2009/07/17
Meeting Minutes
IRC Meeting of Sun Microsystems (StarOffice) with RedFlag2000
Performance Project
Date: 2009/04/03
Time: 15:53– 17:01
Meeting No.:
Agenda:
(3:55:28 PM) FrankL: Moin
(3:59:24 PM) odf-mib: Moin
(3:59:26 PM) Dieter: Hi all :-)
(4:00:07 PM) liheng: Hi,all
(4:00:20 PM) yugq: Hi everyone:)
(4:00:22 PM) cd_oo: Hi all
(4:00:49 PM) arwe: Hi all and good morning :-)
(4:01:01 PM) ycheng1: Hi
(4:01:04 PM) erAck: good morning / afternoon
(4:01:20 PM) erAck: what a beuatiful day..
(4:01:34 PM) liangjun: :)hello all
(4:01:35 PM) Malte: Hi! :)
(4:01:37 PM) Matthias: Hi all
(4:02:10 PM) xiuzhi: hi all
(4:02:12 PM) liheng: Good morning / afternoon!
(4:02:20 PM) FrankL: Hi, all!
(4:02:31 PM) Matthias: Hi liheng, xiuzhi
(4:03:02 PM) xiuzhi: mhu: moin
(4:03:09 PM) Matthias: moin :-)
(4:03:27 PM) liheng: Hi,mhu:)
(4:03:28 PM) liheng: Agenda:
(4:03:28 PM) liheng: 1.Update overall status
(4:03:28 PM) liheng: 2.Exchange of views on the OpenOffice User Survey II: Performance Findings
(4:04:00 PM) xiuzhi: Dieter:moin
(4:04:29 PM) Dieter: Xiuzhi: moin, moin :-)
(4:04:38 PM) ericb2: hi all
(4:04:56 PM) os_ooo: Hi all!
(4:05:06 PM) Dieter: Liheng: Hi!
(4:05:23 PM) liheng: Hi, Dieter:)
(4:07:42 PM) liheng: My status: I have updated a part of benchmark-system code in SVN, benchmark-system/src/benchmark and for Web tools and CRT tools code I need to clear some code that can't use for opensource.
(4:08:40 PM) liheng: And I got the IP for Web Server, it is 219.239.158.91
(4:08:41 PM) Malte: liheng: I have some questions to that
(4:08:53 PM) Malte: There is code which belongs in the OOo repository
(4:09:05 PM) Malte: because you need it for building OOo with benchmark info
(4:09:23 PM) liheng: Malte:ok please
(4:09:27 PM) Malte: So why is SAL/RTL in the performance repository now?
(4:09:28 PM) Malte: Needs to be in the original OOo SAL/RTL module
(4:09:46 PM) Malte: So it would be better to have them in an DEV300 CWS
(4:12:20 PM) liheng: I can't build all code on DEV300 now, the code has been posted into performance repository is the part I just pass build on DEV300
(4:13:05 PM) Malte: Well... on which code base are you working then?
(4:13:34 PM) liheng: I need several day to transform some data-type to OOo sal_XXX
(4:13:39 PM) os_ooo: Status in Writer load/save: The cws os128 with the changes in SfxItemPropertySet, handling of URLs while saving is now being rebased. The access to properties is not faster than before but it is more flexible as it doesn't require alphabetical sorting anymore. Storing documents with links is a lot faster than before. Please have a look at the updated wiki page (thanks to Bjoern Michaelsen)...
(4:13:41 PM) os_ooo: ...at http://wiki.services.openoffice.org/wiki/Performance/WriterInProgress. We found some new areas of improvements like compressing already compressed parts of the document ( jpeg etc. ), access to text frames.
(4:14:15 PM) Malte: liheng: So on which code base are you working ?
(4:15:03 PM) ***Matthias had not had so much time this week; only started an exchange of ideas about some buffered file I/O with tora-japan.
(4:15:29 PM) liheng: It's base Windows based project.:)
(4:15:54 PM) Malte: I mean which OOo code line are you currently using?
(4:16:22 PM) Malte: I assume you already have some code in some OOo version...
(4:16:37 PM) liheng: I have transformed them before I posted them
(4:16:52 PM) liheng: Malte:Yes
(4:17:44 PM) Malte: Well, so it's save to guess it's not OOo 3.x, but some RedOffice code line, and that's the reason why you need to transform something?
(4:17:45 PM) liheng: Malte:Sure,I will create a CWS for native code
(4:19:23 PM) liheng: Malte: Yes, a part is in RedOffice and another is Windows based project. some datatype like __int32 :)
(4:19:29 PM) skotti_: liheng: Please set me (jsk) as QA Rep. for your CWS
(4:19:50 PM) Malte: liheng: I think we should follow up via email...
(4:20:02 PM) Malte: But independently from that:
(4:20:07 PM) liangjun: My status: analyse odf loading and parsing, write some simple applications to compare performance.
(4:20:12 PM) Malte: Please use correct copyright headers
(4:20:23 PM) Malte: not LGPL 2.1 and CVS variables
(4:20:40 PM) liheng: skotti_:Okay!
(4:21:03 PM) liheng: Malte:Where can I find the header?
(4:21:17 PM) liheng: Please send it to me via email
(4:21:18 PM) Malte: In any current OOo source file
(4:21:37 PM) liheng: Malte:I see:)
(4:23:13 PM) cd_oo: Status start up performance: Due to several show stoppers I wasn't able to do much for start up performance. I instrumented additional libraries, e.g. svx, svl and svt - reordering reduces the page fault rate for these libraries about 25%. I now have a pool of 13 optimized libraries. With these optimizations loading libraries on start up is about 15-20% faster. I want to optimize additional...
(4:23:14 PM) tora-japan: my status: (1) Buffered file I/O: I had received a test code from 'mhu' about buffered file I/O implementation in SAL and started looking into it.
(4:23:15 PM) cd_oo: ...libraries and try another reorder strategy. I also want to verify on my test machine the findings of Yuan Cheng that non rebased dlls load faster on cold start up.
(4:23:15 PM) tora-japan: (2) Memory allocation: An attempt of detecting buffer overflow bugs and accessing-to-free()ed-memory-area bugs is ongoing. I am currently compiling X11R6.9.0 on Solaris x86 from scratch since X11 library files embeded in Solaris Express seems to have some problems with debug-version of malloc library called watchmalloc(3MALLOC) http://docs.sun.com/app/docs/doc/819-2243/watchmalloc-3malloc?a=view
(4:24:21 PM) erAck: My status: not much progress from last week, busy with other things. In general continuing work on broadcast areas separation.
(4:25:22 PM) liheng: Malte:I will change web application framework into Glassfish and start service for view diagram and some test result
(4:25:27 PM) skotti_: tora-japan: Do you have a CWS for your changes?
(4:26:08 PM) erAck: tora-japan: for the memory allocation: does that detect things different from Valgrind runs?
(4:26:15 PM) zhangyuwei: odf-mid:Good afternoon, I'm sorry, I lost a file about my code in CWS, and now,I have updated. :-)
(4:26:25 PM) tora-japan: skotti_: I don't think I have such a CWS.
(4:26:30 PM) odf-mib: My status: I'm working on an improvement of the RTL_LOGFILE analysis that includes a claculation of the number of bytes that are loaded or saved per second. And I do some experiments with the performance test to get these numbers for a large number of documents.
(4:26:48 PM) ycheng1: my status : still ongoing for AL in OO3.
(4:27:12 PM) skotti_: tora-japan: Could you create one? Should simplify your work (if it's coming out of experimental stage, that is). I volunteer for QA Rep. here as well
(4:28:07 PM) arwe: Status: Some OOo 3.1 showstoppers forced me to multiple context changes. I started a general comarison of Load/Save performance between OOo, IBM and MS Offices. I am writing a paperdocumenting the measured times and my findings/conclusions. Also added comparing to an OOo version with some added improvements i identified by measurements to check how they improve over the main codeline.
(4:28:39 PM) tora-japan: erAck: I am not sure right now. I will try to study http://valgrind.org/
(4:30:34 PM) Matthias: skotti_: what tora-japan and I are discussing is purely experimental right now; a cws will be created after these experiments have shown something useful.
(4:31:05 PM) Malte: linheng: Sorry for being absent shortly ;) Using Glassfish sounds great! :)
(4:31:15 PM) skotti_: mhu: So i gathered, but if you start coding, get me a CWS very early please.
(4:31:26 PM) peter13j: Hi everyone!
(4:31:32 PM) Matthias: skotti_: sure, will do.
(4:31:47 PM) tora-japan: skotti_: That would be good. I have a ssh-tunnel access to SVN repository, but I don't know if I have a write permission on some modules.
(4:32:11 PM) erAck: tora-japan: because for Valgrind checks you don't need to rebuild X11 ;-)
(4:32:41 PM) Matthias: tora-japan: we'll organize that access, one or the other way.
(4:33:03 PM) odf-mib: zhangyuwei: Unfortunately, all tinderbox build still show errors.
(4:33:07 PM) Matthias: Hi peter13j
(4:33:24 PM) peter13j: Hi mhu
(4:33:55 PM) FrankL: My status: Evaluated the performance part of the OOo User Survey 2009. I have posted it to the performance alias on Wednesday. Please find the presentation here (page 19ff): http://wiki.services.openoffice.org/w/images/8/8d/Renaissance-status-2009-03-26-1.odp I will also blog about this on gullfoss today. Performance logo is in the making. Hope to have it in place until next IRC meeting.
(4:37:00 PM) zhangyuwei: odf-mid:en, Yesterday, I find that i lost a header file, bring on show some errors when building. i had fixed
(4:37:14 PM) skotti_: My Status: "Warming up" for the Performance Project, setup of testing machine (Sun internal) has started. Expecting first results next week.
(4:37:34 PM) tora-japan: erAck: Thank you for the good suggestion: As http://en.wikipedia.org/wiki/Valgrind says, Valgrind checks
(4:37:35 PM) tora-japan: - Reading/writing memory after it has been free'd
(4:37:35 PM) tora-japan: - Reading/writing off the end of malloc'd blocks
(4:37:35 PM) tora-japan: - etc, etc, etc, ...
(4:37:38 PM) tora-japan:
(4:38:34 PM) liheng: FrankL:Thank you, I saw the mail yesterday, and it can help us to define UX Index in benchmark-system
(4:39:09 PM) erAck: tora-japan: Solaris dtrace may also give useful insights. A quite different approach though.
(4:39:45 PM) tora-japan: mhu: I appreciate you.
(4:40:21 PM) tora-japan: erAck: That is interesting.
(4:41:13 PM) odf-mib: zhangyuwei: Okay Then next week, the inderbox status' should be all green.
(4:41:38 PM) odf-mib: inderbox -> tinderbox:-)
(4:41:49 PM) Matthias: tora-japan: I think, the dtrace user manual even has an OOo / StarOffice example for I/O wait tracing :-)
(4:42:44 PM) zhangyuwei: odf-mid: okay ;)
(4:42:55 PM) liheng: FrankL:Do you think the Program start-up is the most important case for OOo performance or the other fatal problems is behind it.
(4:43:19 PM) liheng: FrankL: Do you think the Program start-up is the most
important case for OOo performance or the other fatal problems is behind it?
(4:43:40 PM) tora-japan: mhu: Yes, it does. I coincidently found it. That is wonderful!
(4:43:52 PM) Matthias: :-)
(4:44:33 PM) FrankL: Program start-up is rated as slow, but will happen only very few times a day.
(4:44:50 PM) FrankL: liheng: Thelinked presentation also shos some general information about the usage of OOo. Please also see January and February preenations of the Renaissance project: http://wiki.services.openoffice.org/wiki/Renaissance:Presentations
(4:48:19 PM) FrankL: ...but program start-up is what new users experience first, when they try out OOo. So it becomes more importat for the overall performance experience.
(4:49:26 PM) liheng: FrankL:Lots of people start-up OOo with double click a document in a floder, so in fack, I think some problems is start-up + document loading
(4:49:52 PM) ycheng1: I support liheng's point
(4:51:03 PM) ycheng1: In 'User Feedback program', can the 'double-click opening' action be recorded?
(4:51:41 PM) erAck: FrankL: do we have numbers about how many people use the Quickstarter?
(4:52:20 PM) erAck: FrankL: or leave an instance of OOo running by other means? (which is what I do)
(4:52:31 PM) liheng: erAck:I got a report the Quichstarter of 3.0.x was slower the versions before :(
(4:52:38 PM) FrankL: Yes, I agree. I would call it the warm start. The cold start I was talking about is what users experience when they start OOo the first time.
(4:53:07 PM) erAck: liheng: but once it is running, starting OOo with a document is really fast.
(4:54:09 PM) FrankL: erAck: Hopefully our usage tracking extension (coming wth 3.1) will provide us the data, but I am not sure about that.
(4:55:10 PM) liheng: erAck:Sure, but not faster than early versions
(4:56:54 PM) skotti_: ycheng1: No, loading by double-clicking on the document is currently not recorded
(4:57:07 PM) erAck: liheng: what do you mean by "early versions"?
(4:57:23 PM) liheng: erAck:OOo2.1 2.4
(4:58:35 PM) liheng: The earlier versions of OOo
(5:00:02 PM) erAck: liheng: I didn't refer to earlier versions, just the difference toa cold start (or even warm start)
(5:01:06 PM) liheng: erAck:okay, but I am just finding some reason for it:)
(5:01:07 PM) odf-mib: I have to leave the chat now. Bye!
(5:01:46 PM) liheng: Okay,time up
Performance/Meetings/2009/07/17
Meeting Minutes
IRC Meeting of Sun Microsystems (StarOffice) with RedFlag2000
Performance Project
Date: 2009/04/03 for MM
Time: 15:42– 16:49
Meeting No.:
Agenda:
