Hi Arthur.. Interesting to read your (1st? really?) post, but I am not sure how far I can help...so I'll just take a shot... On Sun, Oct 31, 2010 at 18:36, Arthur Marsh <arthur.marsh@xxxxxxxxxxxxxxxx> wrote: > I have a PII-266 with 384 MiB RAM (maximum capacity for the machine) and > an AMD64 dual core with 4 GiB RAM, both running Debian unstable plus > some packages from experimental and custom kernels. > > A typical load for the PII-266 is KDE 3.5.10 with konversation, icedove, > iceweasel, xmms, lynx, hp-systray, top and aptitude-curses. > > This works with stock Debian kernels and custom kernels up to and > including 2.6.36-git6. Under heavy load, free RAM will hover around 5 > MiB but audio will still play with a very occasional skip and all > applications are responsive. that's free RAM..how about the buffers and page cache size? And how big is your swap size? And is it composed as single partition only in one disk? or...? > With the newer kernels, e.g. 2.6.36-git9,10,11 (all with the deadline > scheduler) get into a swap storm with over 32 MiB RAM free and kswapd0 > taking 10 percent or more of CPU time. Kernel 2.6.36-git15 with the cfq > scheduler managed to keep smaller applications like xmms and shells > running but also had excessive free RAM and kswapd0 taking more than 10 > percent of CPU time. have you tried to play with /proc/sys/vm/swappiness? > On the AMD64 dual core when compiling kernels with CONCURRENCY_LEVEL=4, > I would sometimes get the build process pausing with errors like: > > [ 2880.492025] INFO: task sh:10071 blocked for more than 120 seconds. > [ 2880.493165] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" > disables this message. > [ 2880.494299] sh D ffff8800cfc53780 0 10071 9092 > 0x00000000 > [ 2880.495433] ffff880128b214a0 0000000000000086 ffff880128851b00 > ffff88012b67cba0 > [ 2880.496612] 0000000000013780 0000000000013780 ffff88011969dfd8 > ffff88011969c000 > [ 2880.497765] ffff880128b21798 ffff880128b217a0 ffff880128b214a0 > ffff88011969dfd8 > [ 2880.498925] Call Trace: > [ 2880.500096] [<ffffffff810086fc>] ? __switch_to+0x198/0x284 > [ 2880.501262] [<ffffffff8132d36b>] ? schedule_timeout+0x2d/0xd7 > [ 2880.502418] [<ffffffff8103999f>] ? need_resched+0x1a/0x23 > [ 2880.503564] [<ffffffff8132d13f>] ? schedule+0x5b2/0x5c9 > [ 2880.504739] [<ffffffff8103e36d>] ? get_parent_ip+0x9/0x1b > [ 2880.505883] [<ffffffff8132c9c6>] ? wait_for_common+0x9d/0x10c > [ 2880.507024] [<ffffffff81043d10>] ? default_wake_function+0x0/0xf > [ 2880.508189] [<ffffffff81089999>] ? stop_one_cpu+0x57/0x6e > [ 2880.509288] [<ffffffff81043b23>] ? migration_cpu_stop+0x0/0x2a > [ 2880.510354] [<ffffffff81331951>] ? sub_preempt_count+0x83/0x94 > [ 2880.511426] [<ffffffff8103f880>] ? sched_exec+0xbe/0xd6 > [ 2880.512523] [<ffffffff810fea4b>] ? do_execve+0xd1/0x28e > [ 2880.513582] [<ffffffff81010786>] ? sys_execve+0x3f/0x54 > [ 2880.514616] [<ffffffff81009fdc>] ? stub_execve+0x6c/0xc0 > amarsh04@am64:~$ uname -a > Linux am64 2.6.36-git16 #1 SMP PREEMPT Sun Oct 31 15:41:03 CST 2010 > x86_64 GNU/Linux > > The process was unblocked by logging another session into the AMD64 machine. > > The "task foo blocked for more than 120 seconds" has also occurred on > the PII-266 uniprocessor machine with some of the 2.6.36-git9 or later > kernels. Previously I had not had such a problem on the PII-266 for > about 6 months. > > I can't begin to figure out where this problem was introduced. > Git-bisection doesn't always work as it may take a while for the > symptoms of a swap storm to appear. > > Is there any straightforward way to gather more information before > reporting this kind of issue upstream? Hmmm, I just sense, could be that the scheduler hackers once again play with child-forked-then-run-first? Once I read it is child which runs first, but lately I read it is the parent first... I also think, it might be something related to a patch Fengguang Wu made few weeks ago (IIRC) to reduce latency during high I/O... but I can't recall which patch that is. Try to google... NB: ftrace might help here.... please kindly read the ftrace tutorial posted in lwn.net in few latest editions... -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ