Message boards :
Client :
Client response times + orphans
Message board moderation
Author | Message |
---|---|
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
If you are interested in what is currently going on with the client development you can read the technical details below. Basicly it is about improving the response times between the client application and the BOINC core client so that the core client does not orphan the Blender application (or as it happened a few times, let the Blender application output its debug information into another project\'s workslot). The current response times are ranging from 0.001 to 120 secs with an average of around 20 secs. This is because the controlling application can only handle BOINC messages (pause, shutdown etc) every time it gets a progress update from the controlled application (Blender). Blender only sends progress updates every time it has rendered an entire line - which can take a long time if the image to be rendered is very complex. To improve the response times the idea is to continue execution in the controller even though there are no progress updates. That way the message parsing part of the controller can rely on something else than the progress updates to run (for instance a timer). In the upcomming tests this timer will be set to 0.6 secs, which should make the response times cap at 0.6 sec - it is still possible to get lower response times depending on when the last progress update was done and when the timer last ticked. Furthermore a single condition was found that could potentially kill the controller but leave the controlled application (Blender) running. This is considered a serious bug since the Blender instance will run to completion before exitting, possibly breaking the memory limit set in the BOINC preferences (because other projects can be loaded in). The fix for this will not be included in the first couple of tests, though, since it is important to not vary too many factors at the same time. It only happens in the cases where the built-in error preventer (called the terminator) kills the application. The terminator is designed to shut down everything if the controller looses contact with the controlled application for an extended amount of time, in fact it only shut down the controller. Judging from the debug output returned in previous sessions this happens very very seldomly. ...So you could say that the error preventer caused an error ;) The information needed to find the above bugs was all gathered during the most recent sessions (137-139) as well as some older sessions (131-132, 135). A great thanks goes to the people who participated in those sessions and the mirror owners who made it possible to run the tests at such a large scale. |
![]() Project donor ![]() Send message Joined: 14 Mar 05 Posts: 73 Credit: 25,881 RAC: 1 |
Thanks for the update Janus. You are doing a great job. So, when do you think tose new tests will be run? ![]() when life gives you lemons, make lemonade! |
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
The first internal tests didn\'t show good results. The Blender application got killed way too much. I need to look into this before releasing a series of public tests, as this points towards a so-far unknown issue with the terminator. I\'m afraid the public tests will have to be rescheduled to the next test-timeslot (approx. Jan 25) or even further to the first February timeslot. January is a very busy month for me, so development is a bit slow. Hopefully it will be possible to catch up to speed in the beginning of Feb. The first version that will reach public testing is the Windows one; shortly followed by the linux one. This doesn\'t change the previous schedule much, the only difference is that the client tests that were supposed to be released publicly for the last couple of days were kept in-house due to too many errors. |
![]() Project donor ![]() Send message Joined: 16 Apr 05 Posts: 4 Credit: 3,574 RAC: 0 |
Thanks for all the hard work. Since I know the donations have gone to use (unlike PlanetQuest where I have sent $40 and now they have been quiet for months) I\'ve donated a couple of Euros and will hopefully donate more in the coming weeks. I\'ve also re-attached my systems (3 of them) which 2 of them are to be upgraded to much faster CPUs in a couple of weeks. Also, I even put the homepage as my browsers homepage since I\'m so interested in this project (I\'m a PSP & PC gamer, so animations/models etc grab my attention easily =D ). |
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
Today the client went through 31 recode and evaluation cycles. It is currently running another evaluation cycle and things are looking really well so far. There were a few surprises as well - more about that later. As a consequence of some portability issues the client has been seperated from 2 into 4 different binaries - 2 of which are tiny third party programs (zip and unzip). This will hopefully end the constant battle with library inclusions, overlaps, errors etc. that has taken plase at each previous public release. Also the Windows client is now compiled in cygwin and a GPL\'ed cygwin dll is sent along with the binary to all the users. In other words you should expect more files to be downloaded than usual in the next test - although they will probably each be a bit smaller than the previous client. Here\'s the expected content of the next release: cygwin1.dll - the cygwin library x.zip - Blender release archive x.exe - Controller zip - Archival program used to compress the output before returning it unzip - Archival program used to uncompress the input files and executables sent from the server in.zip - An input file archive containing a part to be rendered Almost all goals for this release have been achieved now, so expect a public test to pop up either tomorrow (or early February if the current test window is missed). |
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
Been through 4 more cycles now. A small problem persists: it doesn\'t respond at all... This seems related to the use of a multimedia timing driver. There\'s a few potential sollutions to this, one will be tested in version 4.24 shortly. [Edit:] 4.24 seems to work better. It will have to run for another 1.5 hours before it is clear whether it works flawlessly. [/Edit] |
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
Been through yet another load of test cycles now. Some of the problems were related to a strange (and difficult do debug) platform issue when compiling windows programs in Cygwin. Version 4.30 has the new code and is currently being tested. |
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
After yet another bunch of cycles it seems the core of this issue has finally been found by one of the participants. Version 4.34 carries a new version of the unzip program as well as updated code. |
![]() Volunteer moderator Project administrator ![]() Send message Joined: 16 Jun 04 Posts: 4574 Credit: 2,100,463 RAC: 8 |
After yet another bunch of cycles it seems the core of this issue has finally been found by one of the participants. Version 4.34 carries a new version of the unzip program as well as updated code. Yup, version 4.34 seems to be the lucky winner of this client development round. So far it has a much lower error-rate than the previous versions (which were at about 60%, this one is more like 5% and the errors do not seem to be related to BURP at all). |
![]() ![]() Send message Joined: 27 Mar 05 Posts: 142 Credit: 3,243 RAC: 0 |
After yet another bunch of cycles it seems the core of this issue has finally been found by one of the participants. Version 4.34 carries a new version of the unzip program as well as updated code. Yes it is working quit well except for those using boinc v5.3.xx. But that is what you get when you don\'t use the recomended version, lol. Well someone has to test them. Those testing boinc v5.3.xx, don\'t forget to report on the boinc site that it has problems with burp so they can fix it. :) Win98SE XP2500+ Boinc v5.8.8 ![]() And God said"Let there be light."But then the program crashed because he was trying to access the 'light' property of a NULL universe pointer. |