This one has the fix for the cache size and high priority problems which have caused some comment on this project.

It's a disaster here. All 5 machines that I installed 6.12.5 on quit running a full complement of CPU programs and refused to run AQUA at all. Reverted to 6.12.4 and they all started behaving properly again except of course for the GPU cache shrinkage bug.

* client: change scheduling policy to allow multithread jobs to coexist with GPU jobs that use significant CPU time.
-- Old: run a MT job only if total CPU usage will be < #CPUs + 1. So if you have some GPU jobs running and their CPU usage sums to < 1, BOINC will run a MT job too. But if CPU usage > 1 BOINC won't run the MT job, and some CPUs will be idle.

Note: to maximize throughput, it might be better to run either GPU jobs or MT jobs, but not both at the same time. However, volunteers don't like it when CPUs are idle. So...
-- New: ignore the CPU usage of GPU jobs in deciding whether to run MT jobs. So we'll run a 4-core MT job (at low priority) even if GPU jobs (which run at normal priority) use > 1 CPU. (Yes, the MT job might run very slow)

There's been quite a big change on this overnight - changeset [trac]changeset:22668[/trac]. You might prefer to defer testing of v6.12.5 after all, and concentrate on v6.12.6 when it comes out - provided you can live with or manage the gpu_active_frac issue in the meantime.

* client: change scheduling policy to allow multithread jobs to coexist with GPU jobs that use significant CPU time.
-- Old: run a MT job only if total CPU usage will be < #CPUs + 1. So if you have some GPU jobs running and their CPU usage sums to < 1, BOINC will run a MT job too. But if CPU usage > 1 BOINC won't run the MT job, and some CPUs will be idle.

Note: to maximize throughput, it might be better to run either GPU jobs or MT jobs, but not both at the same time. However, volunteers don't like it when CPUs are idle. So...
-- New: ignore the CPU usage of GPU jobs in deciding whether to run MT jobs. So we'll run a 4-core MT job (at low priority) even if GPU jobs (which run at normal priority) use > 1 CPU. (Yes, the MT job might run very slow)

Why not just return to the way 6.12.4 and all earlier clients handled this?

Richard

If it helps your BOINC quest, more than happy to run a version for you on my main machine, tell me which one, whatever helps to support you in the latest bug hunt, howsoever that may manifest itself. I am one of those who drifted away from Aqua as such, kind of try it again once in a while but bit half heartedly, due to the effect Aqua MT was having on my GPU crunching. At present I run 6.10.58

I understand the potential hit I might take, and the need to run logs for you, just let me know which logs you want running. It will not be with GPUGRID tasks, as you are aware I have now gone totally AMD on both machines, but if that does not matter to you, no problem. I tend to run MW, Collatz, and DNETC in that order on the GPU, until the AMD app here is deployed.

Regards
Zy

- client, scheduler, RPC protocol:
If the scheduler doesn't have any app versions for resource type X, it includes an element <no_X_apps>1</no_X_apps> in the reply msg (e.g., <no_cpu_apps>1</no_cpu_apps>)
The client parses and stores these flags, and doesn't ask a project for work for a resource if the project doesn't have app versions for it. Apparently I started this change in [19375] (October 2009) and forgot to finish it.

- client: fix "Snooze GPU"

- client: change scheduling policy to allow multithread jobs to coexist with GPU jobs that use significant CPU time. Old: run a MT job only if total CPU usage will be < #CPUs + 1. So if you have some GPU jobs running and their CPU usage sums to < 1, BOINC will run a MT job too. But if CPU usage > 1 BOINC won't run the MT job, and some CPUs will be idle. Note: to maximize throughput, it might be better to run either GPU jobs or MT jobs, but not both at the same time. However, volunteers don't like it when CPUs are idle. So... New: ignore the CPU usage of GPU jobs in deciding whether to run MT jobs. So we'll run a 4-core MT job (at low priority) even if GPU jobs (which run at normal priority) use > 1 CPU. (Yes, the MT job might run very slow)

- manager: don't show resources redundantly on task properties page

- boinccmd: --get_cc_status now prints its result

- client: fix yesterday's checkin [22661] for anonymous platform: if project P is anonymous platform don't request work for resource R from P if there is no app version using R in P/app_info.xml else don't request work for resource R from P if P tells us it has no app versions using R

- MGR: Disable balloons on Linux for the time being.

At this stage, I think I'm going to leave it until Rom gets a v6.12.6 build ready: and then your AMD (ATI hosts) will be useful for double-checking that the work fetch inhibition for devices not supported by the project is working, too (though I see a suggestion that there may need to be server-side updates as well - not sure what the policy on that is at this project).

- MGR: Save window dimensions and position on SIZE and MOVE events to make sure changes are recorded even if the window is minimized (iconized) when exiting BOINC.

- client: account manager RPC:

Additions to request message:
-- <not_started_dur>X</not_started_dur>
-- <in_progress_dur>X</in_progress_dur>
The estimated remaining duration of unstarted and in-progress tasks.

Additions to reply message, within <project>, optional:
<suspend>0|1</suspend> suspend or resume project (overrides local state)
<abort_not_started>0|1</abort_not_started> if set, abort unstarted jobs.

- Add support to detect VirtualBox VM software.

- client and scheduler: a client host may have multiple VM systems installed. TODO: check for VirtualBox on Mac, Linux.

- client: Add support to detect VirtualBox VM software on Mac.

- Port missing minor change from trunk.

- Mac Installer: remove obsolete kludge for generating x86_64 screensaver.

- win_tray: Add logging support to trap crashes (of boinctray).

- client: in account manager requests, include the following per-project attributes: suspended_via_gui, dont_request_more_work, deatch_when_done, ended.

- client: restore code to maintain recent estimated credit.

- client: Add VirtualBox detection for Linux.

- Mac Installer: eliminate wrapper application around install package.

- Mac Installer: permit remote install via SSH using command-line installer.

- client: improve the way credit history is maintained

Old: Maintain list of daily records. When add a new record, delete records older than a month.

Problem: If there's a gap in the record (e.g. because project was down) deleting old records may result in a list that has an entry only for today. Data for the last month is lost.

New: When appropriate, adjust the date of old records rather than deleting them.

- Mac Installer: command-line installs test for existence of files /tmp/nonadminusersok.txt and /tmp/setboincsaver.txt

- MGR: Switch to the tasks tab when a request to execute benchmarks has been requested. The messages tab no longer exists.

- client: fix bug where client would ask project for work for a resource even when the project doesn't have app versions that use the resource.
TODO: there are 2 functions, compute_may_have_work() and dont_fetch(), that do the same thing and both have misleading names. Clean this up.

- Mac Installer: Better handling of custom installer icon.

- MGR Statistics tab: always show last 30 days in all projects (sum) graph.

- Remove BOINC_RCSID tags from more source files.

- client: Copy the AltiVec CPU feature to the p_features field, leave AltiVec in the host model field for backwards compatibility.

- client: implement more scheduler changes that use recent estimated credit (REC) instead of debt.
These changes are enabled by #define USE_REC in work_fetch.h.

If this is commented out (the default) the client uses debt-based scheduling, same as before. TODO: work-fetch policy changes

- client simulator: various fixes:
-- compute idle and wasted fraction based on all processing resources, not just CPU.
-- compute job completion times based on FLOPS, not CPU seconds.
-- compute and use project->no_X_apps etc.

- GUI RPC, client side: add error-checking for CC_STATE::parse(); avoid crash if error or missing item in reply XML.

- client: fix compiler warning.

- vbox: Create a new VirtualBox wrapper using the VirtualBox COM API set. Use the existing wrapper as the baseline. Current wrapper just dumps the existing VMs defined on the machine. More to follow.

- client: work fetch fix: try to maintain GPU work all projects, since we now do round-robin for GPUs as well as CPU. NOTE: this bug was found using the client simulator!

- client simulator: generate REC graph.

- client: do exponential backoff (from 10 min to 24 hours) on account manager RPC failures, rather than always waiting 24 hours.

- client: use <foo>0|1</foo> instead of <foo/> or "" as XML boolean. This distinguishes between "false" and "not defined" (did this in a few places, should do it everywhere).

- client: send host info in account manager request message.

- client: on startup, generate a notice if any jobs require more RAM than is allowed by prefs.

- manager: show RAM limit in task properties.

- api: check return values of functions used in boinc_msg_prefix(), return a (valid) empty string if an error occurred.

- MGR: Use standard menus, even on Windows, for the Simple GUI.

BOINC 6.12.8 released for testing for all platforms

Disclaimer
On development versions of BOINC:
- Expect parts of it to be broken, or in the least to work in a different way than you are used to.
- Expect work failures, deadline misses and losing all your accumulated work in progress, or not getting credit for your work due to unknown and unforeseen circumstances.
- Only use it when you are willing to participate in the Alpha test program and do not mind sending (bug) reports in to the developers on the email list specially assigned for this purpose.


REMINDER TO ALL ALPHA TESTERS:
It's far easier for the developers to fix problems when you send message logs with the appropriate flags set. The main flags are:

<cpu_sched_debug>: problems involving the choice of applications to run.
<work_fetch_debug>: problems involving work fetch (which projects are asked for work, and how much).
<rr_simulation>: problems involving jobs being run in high-priority mode.
<std_debug>: problems involving short term debt.

Use these flags from the cc_config.xml file.

Report any problems you get with it to the Alpha email list. This list needs registration.

Reporting bugs on this forum or any project's forum isn't very productive as the developers simply do not have time to scour through all forums looking for posts with your bug reports in them. While it may happen that we - other volunteers - forward your problem to them, history has shown it is quite difficult being the middle man in this when more information is needed from you. Hence why it's better for all concerned that you report the problem directly to the developers, who will come back directly to you when they need more information or better logs or to tell you that they've checked in a fix for the specific problem.

The actual change log was taken from the Alpha email list and the Trac Timeline repository changes check-ins. This disclaimer was written by the volunteer moderators of this forum and approved of by the BOINC development team.

Change Log:


lib: The custom Mac backtrace code doesn't work on Intel Macs running under OS 10.5.x, so use newer backtrace() and backtrace_symbols() APIs under OS 10.5.x, with explanatory text suggesting alternatives. (Checked into 6.12.7 tag.)
client: fix major bug in job scheduling

Old: scheduling has 2 phases:
1) computing a list of jobs to run (a vector of RESULT*); this is done infrequently
2) enforcing the schedule; this is done more frequently

Problem:

when we enforce the schedule, the RESULTs in the jobs-to-run list may not still be runnable, and in fact they may not still exist (dangling pointer).

New: combine the 2 phases

client, acct manager protocol: allow <no_cpu>, <no_cuda> and <no_ati> bools within <account> in reply message. They suppress work fetch for that resource type from that project.
Mac lib: use newer backtrace() APIs under OS 10.5.x on Intel Macs.
Mac Installer: Updates for OS 10.6.5 compatibility.
client: fix scheduling bug. The round-robin simulation wasn't handling multithread jobs correctly.
For example, given two 3-CPU jobs, it would model running them together on a 4-CPU host. This doesn't correspond with the CPU scheduler, which runs only 1 at a time. So the simulator would say that there are no idle CPUs when in fact there are, and no new CPU jobs would be fetched.

client: fix typo that broke communication with account managers.
API: compile fix for Win.
client: slightly change detach scheme for account managers.

6.12.9 for the Mac are on the download page. Not too sure what issue(s) they address though.

- manager: fix compile warnings; from Steffen Moller

- client: fix scheduling bug when a job has fraction_done = 1 but it's not finished yet. ACTIVE_TASK::est_dur() was returning 0 when it should have returned elapsed_time.

- client: fix bug where "reread config file" didn't update the max log file sizes

- client: fix scheduling problem:
Old: enforce_schedule() won't run an active job if its working set size exceeds remaining available RAM.
Problem: there may be a lot of similar jobs. The client starts one, finds that its working set is too large, starts the second, and so on.
Solution: if J is an unstarted job, and there are started jobs using the same app version, consider J's working set size to be the largest of the working sets of those jobs.

- client: fix an apparent bug that could oversaturate the CPUs with single-thread jobs

- client: work fetch: if the chosen project is currently uploading a file, and an upload started in the last 5 min, don't fetch work from it. The goal is to merge the 2 scheduler RPCs (fetch work, report completed taskS) into a single RPC. Note: this may result in idleness in some cases.

- client: if detect that system clock has been set backwards (either at startup or during execution) reset a number of "wait until X" variables; otherwise we might wait years to contact a project, restart a file xfer, etc. Notes: there is no problem setting clocks forward; things just happen prematurely, some variables (e.g. task deadlines) are not reset, because it's not clear what to set them to

- client: improve text of "message from server" notices

- manager: change "add account manager" to "use account manager". "Add" is confusing, because you can't add multiple account
managers like you add projects.

- client: recognize a few new ATI GPU models from [P3D] Crashtest

- client: tweak to ATI CPU recognition from [P3D] Crashtest

- unix build: remove hardwired -ldl

- client: show a given system notice at most once a week

- Mac: Fix symbol table generation

- client: use std::deque instead of std::vector for RR sim's pending-job lists. Erasing head of vector is slow.

- lib: allow GPU peak FLOPS to be specified in XML (for simulator)

- client: old work fetch policy: projects may need enough jobs for all device instances, not just resource_share*ninst. E.g. a project that has only CPU jobs in a CPU/GPU client

- client: with REC scheduling, don't ask for work for secondary resources if project has negative priority

- client: in RR sim, make sure we saturate devices if possible. Otherwise we may report a shortfall incorrectly

- client: set PROJECT::last_upload_start when a job completes, NOT when its first upload starts. This prevents a scheduler RPC from happening between the two. This was a bug in [22841]

- client: tweak [22842] so that system clock changes of < 60 seconds are ignored

- MGR: Embed the BOINC skin in the executable using XPM files.

- single-job submission system: don't hardwire i686 architecture; make it work on any Linux system. From Steffen Moller.

- MGR: Disable the background repaint events for the SGUI Messages and SGUI Preferences dialog. Changing the font color wasn't as easy as I had hoped and the black background for the skin doesn't do any good when the text itself is black as well.

- scheduler/client/API: pass user ID and team ID from scheduler to client, and from client to app (host ID is already passed).

- MGR: re-enable background painting and just change the default background color for the new default skin to light gray.

- unix build system: generate_svn_version
generate version string from plain git repositories, too
define SVN_REPOSITORY (URL) and SVN_REVISION (numerical) separately so these could be used in proprocessor directives
make the Manager build on FC11

- client: fix job scheduling bug.
Scenario: we have 2 CPUs and 1GB RAM Project A has jobs that use 1GB. Project B has small-mem jobs. Currently, we ignore RAM usage in building the runnable job list, so we might pick 2 jobs from project A. enforce_schedule() (which checks RAM usage) will run just one of them, and the other CPU will be idle.
Solution: keep track of RAM usage while building the runnable job list.

- client: code cleanup in CPU scheduling; eliminate global variable

- Mac: update release script to remove install of separate BOINC skin.

Mac: update release script to remove install of separate BOINC skin.

- simulator work

- client: eliminate unnecessary CPU reschedules. Currently we do a reschedule any time a job checkpoints, in case there's a job that has finished a time slice but hasn't checkpointed yet. Instead: flag such jobs, and trigger a reschedule on checkpoint only for flagged jobs.

- client: fix instability in job scheduling that happens if a job's estimated completion time in RR sim is close to its deadline. It can alternate between making and missing deadline, causing the scheduler to alternate rapidly between jobs. Solution: if RR sim has marked a job as deadline miss any time in the last (CPU scheduling period), treat it as a deadline miss.

- client: if we're using an acct mgr, and we're attached to a project manually (not via acct mgr) and the acct mgr tells us to detach from it, don't.

- MGR: Fix advanced view button is Simple GUI. (Button Text is Verdana 8pt. in MS Paint)

- MGR: Save window dimensions and position on SIZE and MOVE events to make sure changes are recorded even if the window is minimized (iconized) when exiting BOINC.

- client: accept <options/> and <log_flags/> in cc_config.xml

- client: reset log flags (as well as options) when rereading cc_config.xml

- Mac: Add info about dual-GPU Powerbooks to ReadMe files.

- Mac: Localize items in Mac's BOINC menu.

- MGR: Use branded short application name in Help tooltips instead of always using "BOINC".

- Mac: Eliminate unused source files Localization.cpp, .h from Mac project.

- client: make "wrong URL" message not be a notice

- WINSETUP: Avoid a 0xc0000135 error by giving the boinc_master and boinc_project membership in the 'Users' group. Dynamic linking to nvcuda.dll was failing because the project applications were being denied access to DLLs in the system directory.

- MGR: Enable the volunteer configurable 'shutdown core client' feature on exit for Linux. It was orginally supposed to be enabled when the balloon features were enabled, but got lost in ifdef hell.(From "Lee")

- client/manager: show venue in project properties

- manager: msg tweak

- client: don't crash if omit cmdline port#. Fixes #1040

- MGR: Fix a bug that is several years old, appears to have been introduced during the Unicode conversion phase for Windows. Passing a wide character string into an narrow character function causes a NULL reference exception. TODO: There is still an issue if the manager attempts to write to stderr when the BOINC Manager log files are redirected to a directory in which the user doesn't have access too.