|
Strange cpu values for task rcu_preempt
After some time of computing I get some strange %cpu and cpu time values for task rcu_preempt, PID 12:
:: Code ::
ps axo pcpu,comm,pid,user | sort -nr | head -n 3 373415872 rcu_preempt 12 root 58.8 einsteinbinary_ 1932 boinc 58.6 einsteinbinary_ 1934 boinc ps axo time,comm,pid,user | sort -nr | head -n 3 213503-23:34:55 rcu_preempt 12 root 00:48:09 einsteinbinary_ 1932 boinc 00:48:00 einsteinbinary_ 1934 boinc When using liquorix kernel 3.6 it was rcuc/0, PID 9, who showed similar weird values. Could this issue be related to the BFS task scheduler since I never noticed this when using a Debian kernel? My system: :: Code ::
inxi -bAxx System: Host: debian Kernel: 3.7.0-1.dmz.1-liquorix-amd64 x86_64 (64 bit, gcc: 4.7.2) Desktop: Xfce 4.10.0git-1919aa9 (Gtk 2.24.13) dm: lightdm Distro: Debian GNU/Linux 7.0 Machine: System: Acer (portable) product: TravelMate 4530 version: V1.11 serial: LXTPN0C01491707B691601 Mobo: Acer model: Elgon serial: LXTPN0C01491707B691601 Bios: Acer version: V1.11 date: 08/06/2008 Chassis: type: 10 serial: LXTPN0C01491707B691601 CPU: Dual core AMD Athlon X2 QL-64 (-MCP-) clocked at 2100.00 MHz Graphics: Card: Advanced Micro Devices [AMD] nee ATI RS780M/RS780MN [Mobility Radeon HD 3200 Graphics] bus-ID: 01:05.0 chip-ID: 1002:9612 X.org: 1.13.99.901 drivers: ati,radeon tty size: 156x47 Advanced Data: N/A for root Audio: Card-1: Advanced Micro Devices [AMD] nee ATI RS780 HDMI Audio [Radeon HD 3000-3300 Series] driver: snd_hda_intel bus-ID: 01:05.1 chip-ID: 1002:960f Card-2: Advanced Micro Devices [AMD] nee ATI SBx00 Azalia (Intel HDA) driver: snd_hda_intel bus-ID: 00:14.2 chip-ID: 1002:4383 Sound: Advanced Linux Sound Architecture ver: k3.7.0-1.dmz.1-liquorix-amd64 Network: Card-1: Broadcom NetXtreme BCM5764M Gigabit Ethernet PCIe driver: tg3 ver: 3.125 bus-ID: 06:00.0 chip-ID: 14e4:1684 Card-2: Ralink RT2790 Wireless 802.11n 1T/2R PCIe driver: rt2800pci ver: 2.3.0 bus-ID: 05:00.0 chip-ID: 1814:0781 Drives: HDD Total Size: 250.1GB (25.5% used) Info: Processes: 151 Uptime: 28 min Memory: 1074.8/3698.0MB Runlevel: 5 Gcc sys: 4.7.2 Client: Shell (bash 4.2.37) inxi: 1.8.24 Back to top |
|||||
|
Could be, yes. BFS accounts for cpu time much differently than CFS. I'll keep looking a this, it could also be that BFS performs worse with rcu_preempt, and your output is a symptom.
Back to top |
|||||
|
New kernel is out without rcu_boost, a feature that's known to conflict with BFS. Can you see if you get the same anomalies?
Back to top |
|||||
|
I installed the new kernel.
After 3 hours of computing, no abnormal values. Looks good but to be sure we'll have to wait until tomorrow. Back to top |
|||||
|
One day later, again the same anomaly:
:: Code ::
ps axo pcpu,comm,pid,user | sort -nr | head -n 2 28343414 rcu_preempt 9 root 67.5 einsteinbinary_ 1714 boinc ps axo time,comm,pid,user | sort -nr | head -n 2 213503-23:37:07 rcu_preempt 9 root 12:11:50 einsteinbinary_ 1714 boinc Should I distrust BFS? Back to top |
|||||
|
Can you try 1.dmz.4? I the time accounting fix in BFS v426 actually created more bugs than it fixed. Thats been reverted in the newest package.
Back to top |
|||||
|
I already tried 1.dmz.4 and almost immediately got a
similar abnormal rcu_preempt %cpu value. Running Debian's linux-image-3.7-trunk-amd64, the task manager doesn't show any weird value. Back to top |
|||||
|
I installed linux-image-3.7.0-1.dmz.5 and get:
:: Code ::
ps axo pcpu,comm,pid,user | sort -nr | head -n 4 215365543 jbd2/sda5-8 1353 root 202384220 rcu_preempt 9 root 56.3 einstein_S6LV1_ 1711 boinc 55.6 hsgamma_FGRP2_0 1713 boinc ps axo time,comm,pid,user | sort -nr | head -n 4 213503-23:34:40 rcu_preempt 9 root 213503-23:34:33 jbd2/sda5-8 1353 root 00:16:36 einstein_S6LV1_ 1711 boinc 00:16:26 hsgamma_FGRP2_0 1713 boinc Another BFS kernel which I have tried, linux-image-3.7.1-pf_1_amd64.deb doesn't show these abnormal values. Back to top |
|||||
|
Interesting, I'll take a look at the pf kernel and find out what's different.
Back to top |
|||||
|
Well, I can't get a straight answer out of Con, but he doesn't think that this is completely BFS's fault. In that case, the 3.7.0-1.dmz.6 and newer now run CFS instead. I don't think it's a good idea to run a scheduler with a known timer bug that can cause all sorts of random unexplainable behavior.
Please let me know if you still get that error. The sudden changes in liquorix were caused by a third confirmation on Con's own blog page here: ck-hack.blogspot.com/2012/12/37-ck1-bfs-426-for-linux-37.html?showComment=1356960043310#c4842952502245392347, which are exactly what you've described here. Back to top |
|||||
|
All times are GMT - 8 Hours
|
|||||