Message boards :
Theory Application :
On old PC's all WU runs almost exactly 18 hours: is it strange or normal? On modern PC's runtime is very different...
Message board moderation
Author | Message |
---|---|
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
On old PC's all WU runs almost exactly 18 hours: is it strange or normal? On modern PC's runtime is very different... |
Send message Joined: 24 Oct 04 Posts: 1114 Credit: 49,501,728 RAC: 4,157 |
Normal The main thing is we prefer them to be Valid tasks. |
Send message Joined: 15 Jun 08 Posts: 2386 Credit: 222,928,861 RAC: 137,663 |
Theory tasks are designed for a runtime of 12 h. After that 12 h running subtasks will be given the time to finish but no new subtask will be started. 18 h is a hardwired limit checked by a watchdog to prevent crashed VMs from running forever. The watchdog shuts down the VM even if a subtask is not yet finished. This forced shutdown mostly affects sherpa jobs as they tend to have by far the longest runtimes. As MAGIC already wrote the tasks are marked as valid (from the BOINC perspective) as it's not a volunteer's fault to shut down the VM. |
Send message Joined: 1 Feb 06 Posts: 66 Credit: 9,723 RAC: 0 |
Kucky you that some results returned valid... https://lhcathome.cern.ch/lhcathome/result.php?resultid=214740242 After hours, by somehow of miracle, job was cancelled and sent to "heaven" with 0 credits & results. Why? None knows. As soon as I get Theory tasks, I cancel them. Atlas more or less works fine... |
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
MAGIC, computezrmle, thank you for your answers. I wonder, does old PC returns any useful result to the Theory Project if VM do not finish normally? And is it true that I waste about 6 hours of processor’s time for nothing with each WU on old PC? Can I do something to do work more effectively? |
Send message Joined: 15 Jun 08 Posts: 2386 Credit: 222,928,861 RAC: 137,663 |
Good subtasks (= jobs) show the following pattern in your logfiles: 2019-01-28 19:53:45 (4064): Guest Log: [INFO] New Job Starting in slotx 2019-01-28 19:53:45 (4064): Guest Log: [INFO] Condor JobID: 487756.33 in slotx 2019-01-28 20:01:36 (4064): Guest Log: [INFO] Job finished in slotx with 0. In case the watchdog shuts down the VM only the last unfinished subtask is lost but will be rescheduled. To get a better average performance (Theory Simulation) you may run several tasks in parallel as 1-core setup instead of a single task as multicore. This avoids idle cores between 12 h and 18 h runtime. Unfortunately ATLAS interprets the websettings in a different way (wrong in my eyes) thus reducing the #cores will also reduce the #tasks your client can download. This especially affects modern computers with lots of cores. You can work around the negative impact if you use an app_config.xml or if you run extra BOINC clients. |
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
computezrmle, thank you. I'll find out this way. Thanks again :-) |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
...... or use a separate venue for your slower computers. As computezrmle said: for Theory it's the best when you use only single core tasks (if you have enough memory). |
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
Crystal Pellet, thank you. And… Now I see, it is true: single core tasks is one and a half times more effective than multicore :-) |
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
And now it becomes very interesting: is the same true for ATLAS? I runs ATLAS on modern PCs and I want to do it the best way... |
Send message Joined: 13 Apr 18 Posts: 443 Credit: 8,438,885 RAC: 0 |
Yes, it is true for ATLAS too but ATLAS tasks take much more RAM than Theory so if the PC doesn't have enough RAM then you might not be able to run a single core ATLAS task on all cores concurrently. In such cases you might want to run 2-core ATLAS tasks. For ATLAS VBox, RAM = 3000 MB + 900 MB * ncpus For Theory VBox, RAM = 630 MB + 100 MB * ncpus ncpus = number of cpus assigned to the task |
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
bronco, thank you :-) I'll experiment and share my experience here. |
Send message Joined: 18 Nov 17 Posts: 119 Credit: 51,287,231 RAC: 20,565 |
Yes, it is true for ATLAS too but ATLAS tasks take much more RAM than Theory so if the PC doesn't have enough RAM then you might not be able to run a single core ATLAS task on all cores concurrently. In such cases you might want to run 2-core ATLAS tasks. I discovered a very strange situation... If I run 4-core ATLAS tasks, I can get only 4 tasks... If I run 3-core ATLAS tasks, I can get only 3 tasks... If I run 2-core ATLAS tasks, I can get only 2 tasks... Why so? How can I run, for example, 12 2-core ATLAS tasks at 24-core PC? |
©2024 CERN