Message boards :
Sixtrack Application :
Very low CPU-usage on Windows with SixTrack tasks
Message board moderation
Author | Message |
---|---|
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
On my Windows7 machine SixTrack tasks use very low CPU after the start (5%). This can go up up to 55 minutes almost idling and then suddenly the tasks are using a full core. This is happening with the 2 kinds of SixTrack applications I got fo far: pni and sse2. What's happening or what's wrong with my machine. https://lhcathome.cern.ch/lhcathome/results.php?hostid=10362384&offset=0&show_names=0&state=4&appid= On the same hardware the Linux tasks are doing well. |
Send message Joined: 21 Jun 10 Posts: 40 Credit: 10,608,629 RAC: 10,627 |
I noticed something similar on my Windows 10 machine. When a Sixtrack task started, it used about 40% of a thread until the task was ~7% complete (~3 minutes). Then the % complete jumped back to almost 0% and it started using 100% of a thread. |
Send message Joined: 28 Sep 04 Posts: 674 Credit: 43,161,705 RAC: 16,020 |
The low CPU usage in the beginning of sixtrack is normal. While the application uses little CPU the load is on Windows system programs (conhost and csrss). I reported this earlier, you see it here https://lhcathome.cern.ch/lhcathome/forum_thread.php?id=3998#27788 |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
Thanks Harri for the link. Before posting I searched for 'low cpu' and not for 'High non-Boinc CPU', so I didn't find something useful. However I had tasks up to 55 minutes low application usage on a 30-core Windows-VM and not on a 14-core Linux-VM or on a 8-core Win7 host. During that 'long' startup phase (BOINC CPU-setting 100%) OS task monitor was much lower then the maximum - about 40-50% maybe. I'll see and try to monitor that behaviour - else I'll detach LHC@home from that machine - too much CPU-waste. Edit: If I could get SixTrack tasks. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
I got 4 SixTrack's on that 30 core machine and monitored the low cpu behaviour of the sixtrack_win32_4517_sse2.exe. During the first 12 minutes almost no CPU-usage, but saw csrss.exe using about 20% of 1 core and a conhost.exe (child of csrss) using about the same. After those 12 minutes the 2 above mentioned processes sleep and sixtrack_win32_4517_sse2.exe started to use a whole core. Maybe running more tasks concurrently may slow down the init-phase much more. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
Maybe running more tasks concurrently may slow down the init-phase much more. I did a new test with 21 concurrent tasks. The single csrss.exe used about 2 cores of CPU-time and the 21 conhost.exe's together about 3 cores. The 'idle' period of the 21 SixTrack.exe's lasted almost 43 minutes. Note: I'm running this BOINC in a Virtual Machine (Win7) (VMWare) - maybe related. I'll later test on a 8-core Win7 host directly and see how the SixTrack behave there. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
I'll later test on a 8-core Win7 host directly and see how the SixTrack behave there. Same behaviour with 8 tasks running, but 'idle' time of the SixTrack.exe's was much shorter. 'Only' a bit over 3 minutes, where csrss.exe used about 1 core and the 8 conhost.exe's together about 2 cores. Have to add that on this machine, I only got the 64bit version sixtrack_win64_4517_sse2.exe. Can the server not been told not to send 32bit tasks to a 64bit machine? |
Send message Joined: 15 Jun 08 Posts: 2386 Credit: 222,991,823 RAC: 136,421 |
Can the server not been told not to send 32bit tasks to a 64bit machine? See: http://boinc.berkeley.edu/wiki/client_configuration You may set the following option in your cc_config.xml: <no_alt_platform>0|1</no_alt_platform> If enabled, the client will run applications only for its primary platform. For example, a Win64 machine will run only Win64 apps, and not Win32. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
You may set the following option in your cc_config.xml: Yeah, I know, but this setting is BOINC-wide and you will also not get tasks from projects only providing 32bit applications. |
Send message Joined: 12 Jul 11 Posts: 857 Credit: 1,619,050 RAC: 0 |
Dear All, thanks for your feedback. We had some previous discussion at https://lhcathome.cern.ch/lhcathome/forum_thread.php?id=3998#27788 I did try some tests on my Windows 7 box (4 threads) and indeed Windows was very slow to start SixTrack. SixTrack has three phases, pre-processing, tracking, and post-processing. The tracking is almost all CPU and may last from 0 seconds (unstable beam) or for several hours. The pre-processing should take less than 1 minute, but involves a lot of file opens (more than 60) and some I/O. My own tests on 4 threads do show very slow pre-processing but not as bad as described by you. I think the Windows task scheduling is very poor . There may also be locking issues on a multi-core system, especially as we are running an old executable from Windows XP from more than 2 years ago! (This is a very nasty problem. I cannot build the BOINC API libraries on Windows 7 or 10 and CERN has abolished XP. I want to use gfortran and gcc as the ifort compiler for Windows is not numerically compatible. I used to use cygwin but I can't install it anymore becuase of some Tex very large files never completing (known problem.). I am trying MinGW (we do NOT use Visual Studio) but the script seems out of date. I have also tried xcompile but gfortran seems to be installed in the wrong place.... I really must get some help from my Windows support at CERN or from other projects.) Eric. |
Send message Joined: 27 Sep 08 Posts: 798 Credit: 644,814,926 RAC: 229,440 |
Hi Eric, maybe you could talk to David from the BOINC team to get some help with building the BOINC libraries? He seems active with the ATLAS project team. |
Send message Joined: 16 Apr 14 Posts: 2 Credit: 86,718 RAC: 0 |
Have you considered MSYS2? It is more up-to-date than MSYS. I have used it with NetBeans as an IDE on Windows 10 and it seems to work OK. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
This time I got a bunch SixTrack tasks on the same Windows machine , where I reported the low cpu-usage, but now I got only pni-tasks (32- and 64bits) and now the init-phase was only 12 seconds. In OP-mentioned tasks weren't pni and those longer running tasks had a CPU efficiency of 88%. Today the CPU-efficiency of the longer running pni-tasks was 99,6%. |
Send message Joined: 12 Jul 11 Posts: 857 Credit: 1,619,050 RAC: 0 |
Thanks Toby, I am getting some help from Einstein@home and I have now managed to build two out of three libs I need. I am still missing the zip. Eric. |
Send message Joined: 12 Jul 11 Posts: 857 Credit: 1,619,050 RAC: 0 |
Good point, I'll check that I am indeed using msys2. Eric. |
Send message Joined: 12 Jul 11 Posts: 857 Credit: 1,619,050 RAC: 0 |
Thanks for this feedback. If you have SSE/PNI you should indeed get the sse2 or sse3 executable. I am working on all this. Eric. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
The issue seems to be batch specific. Today tasks have the same issue: low cpu usage of the sixtrack executable and 1 csrss.exe some cpu depending how many tasks are running and for each task 1 conhost.exe a bit of cpu. 1 workunit (the first returned one is mine): https://lhcathome.cern.ch/lhcathome/workunit.php?wuid=52870548 You see the big difference between run time and cpu time. There was cpu left on that system and BOINC is allowed to use 100%. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
I digged a bit more into these tasks on Windows with very low CPU-usage at the start. When starting a single task about 11 minutes at the start the CPU-usage is between 10% and 20%. When the real processing starts a lot of tasks thereafter are ready within 2 minutes. When starting more tasks at once these low CPU-time increases up to over one hour when all 30 cores are used. In my former post I already mentioned the issue seems to be batch specific. All tasks starting their names with w-c8_lhc2016 and w-c9_lhc2016 don't suffer from these long idling timeouts during the first phase. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
Any progress? With CPU 100% allowed on a Windows 7 machine: LHC@home w4_hllhc10_round_a4b5_000v665_w4__51__s__62.31_60.32__2_4__5__22.5_1_sixvf_boinc11730_1 451.07 SixTrack (pni) elapsed 00:35:08 CPU 00:00:07 After 36 minutes real processing started. |
Send message Joined: 14 Jan 10 Posts: 1268 Credit: 8,421,616 RAC: 2,139 |
Show it with a picture. 30 threads, first started 15 Sixtracks and after half an hour the other 15. Taking the picture one task was uploaded and acknowledged https://lhcathome.cern.ch/lhcathome/result.php?resultid=147624816 In the column v2 you see the elapsed time and (used cpu time). So real processing started after 28 minutes. |
©2024 CERN