Message boards : Graphics cards (GPUs) : MDIO ERROR: cannot open file "restart.coor"
| Author | Message |
|---|---|
|
As above. After having successfully crunched 2 wus, the following 2 were aborted with this kind of log: | |
ID: 12738 | Rating: 0 | rate:
 /   Reply
Quote
| |
|
MarkJ  Volunteer moderator Volunteer tester Send message Joined: 24 Dec 08 Posts: 738 Credit: 200,909,904 RAC: 0 Level  Scientific publications       | |
As above. After having successfully crunched 2 wus, the following 2 were aborted with this kind of log: We all get the MDIO error message when work units first start. Its the last line thats the error message. I notice you have 6.6.28 which as far as I know doesn't support the later cuda plan classes (cuda23). This problem may go away if you use a more up to date version of BOINC. GPUgrid recommends 6.10.3 (which is an alpha-test version, but seems pretty stable). ____________ BOINC blog | |
| ID: 12741 | Rating: 0 | rate:
/ Reply
Quote
| |
|
Richard Haselgrove Send message Joined: 11 Jul 09 Posts: 1620 Credit: 8,904,489,131 RAC: 19,429,070 Level  Scientific publications | |
I notice you have 6.6.28 which as far as I know doesn't support the later cuda plan classes (cuda23). This problem may go away if you use a more up to date version of BOINC. GPUgrid recommends 6.10.3 (which is an alpha-test version, but seems pretty stable). I'm pretty sure this is another urban myth. I'm still running v6.6.37 for WinXP, but I see I have <app_version> <app_name>acemd</app_name> <version_num>667</version_num> <platform>windows_intelx86</platform> <avg_ncpus>0.129308</avg_ncpus> <max_ncpus>0.129308</max_ncpus> <flops>61084800000.000000</flops> <plan_class>cuda23</plan_class> <api_version>6.7.0</api_version> <file_ref> <file_name>acemd_6.67_windows_intelx86__cuda23.exe</file_name> <main_program/> </file_ref> ... in my client_state.xml - that must be from last time I ran GPUGrid, been running SETI for a while until they run out of data ;-) Just tried getting a new task: was allocated 6.71_cuda.exe, plan class CUDA and v2.2 DLLs - but that's a project decision, not a client decision. | |
| ID: 12743 | Rating: 0 | rate:
/ Reply
Quote
| |
|
MarkJ Volunteer moderator Volunteer tester Send message Joined: 24 Dec 08 Posts: 738 Credit: 200,909,904 RAC: 0 Level Scientific publications | |
I notice you have 6.6.28 which as far as I know doesn't support the later cuda plan classes (cuda23). This problem may go away if you use a more up to date version of BOINC. GPUgrid recommends 6.10.3 (which is an alpha-test version, but seems pretty stable). Well his core client was listed as 6.6.28 (see top message in thread). However regarding your version of the science app, that depends on what nvidia drivers you have installed. It will use cuda 2.3 if your drivers support it and as far as I understand it the cuda23 plan class. But then you are running 6.6.37 which is more up to date than his version. It might just be that his 6.6.28 core client isn't clearing vram properly. An upgrade (depends on version) should fix that. ____________ BOINC blog | |
| ID: 12744 | Rating: 0 | rate:
/ Reply
Quote
| |
|
Richard Haselgrove Send message Joined: 11 Jul 09 Posts: 1620 Credit: 8,904,489,131 RAC: 19,429,070 Level Scientific publications | |
Well his core client was listed as 6.6.28 (see top message in thread). The actual plan classes (cuda or cuda23) are implemented server-side, and the client - having reported hardware and driver versions - accepts what it's given. That's how I come to have plan class 'MT' in my AQUA project, and how I got cuda23 into GPUGrid. I'm pretty sure that there weren't any changes to the plan class mechanism between .28 and .37 - and I was following that particular phase of the development pretty closely, as you know from boinc_alpha! My drivers (on all three machines) are 190.38, installed no later than 22 July and running unchanged since that date. I tested the 2.3 DLLs here manually on 31 July (message 11550), and tested the cuda23 plan class on 20 August (message 11979). Ah, so that's where the entry in client_state came from! All in all, I'm pretty certain that cuda23 works with any BOINC 6.6.xx, and the only immediate reason to upgrade to v6.10.xx is to gain ATI support - though obviously they're not going to back-port many more bugfixes into v6.6, and we'll all have to switch to v6.10 eventually if we want to go on picking holes in the development process. Edit: in that 20 August thread, GDF said that in order to use the cuda23 plan class, The client needs to be able to report the driver version. I think that's true of all v6.6.xx clients for Windows, and certainly not true of v6.4.x Macs are significantly different - none of this discussion applies to them. | |
| ID: 12746 | Rating: 0 | rate:
/ Reply
Quote
| |
|
MarkJ Volunteer moderator Volunteer tester Send message Joined: 24 Dec 08 Posts: 738 Credit: 200,909,904 RAC: 0 Level Scientific publications | |
|
@ Richard | |
| ID: 12751 | Rating: 0 | rate:
/ Reply
Quote
| |
Message boards : Graphics cards (GPUs) : MDIO ERROR: cannot open file "restart.coor"
