Re: Commit fcb6a15c2e7e (intel_pstate: Take core C0 time into account for core busy calculation) sucks rocks
From: Dirk Brandewie
Date: Thu Feb 20 2014 - 09:56:41 EST
On 02/19/2014 04:51 PM, Greg KH wrote:
On Wed, Feb 19, 2014 at 04:35:37PM -0800, Greg KH wrote:
On Wed, Feb 19, 2014 at 04:03:40PM -0800, Dirk Brandewie wrote:
On 02/19/2014 02:47 PM, Greg KH wrote:
Hi Dirk,
I've been having some huge slowdowns on my box building kernels, and I
took the time to bisect it down to commit
fcb6a15c2e7e76d493e6f91ea889ab40e1c643a4 (intel_pstate: Take core C0
time into account for core busy calculation). With that patch reverted
on Linus's current tree, my build speeds are back up to the normal rate.
The difference is huge, 2 minutes to do a kernel build with that patch
reverted, 8-10 minutes with it applied! With all of the stable kernel
builds and other trees, this is a huge problem for my workload (all I do
is kernel builds it seems...)
I see some patches you marked as "fixes" that you sent to Rafael, do you
want me to test any of those? How am I the only one seeing this
problem, do you need my cpu information or anything else?
Can you give me a description of you build system? CPU, number of sockets,
The last processor in /proc/cpuinfo:
processor : 7
vendor_id : GenuineIntel
cpu family : 6
model : 60
model name : Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz
stepping : 3
microcode : 0x9
cpu MHz : 3556.464
cache size : 8192 KB
physical id : 0
siblings : 8
core id : 3
cpu cores : 4
apicid : 7
initial apicid : 7
fpu : yes
fpu_exception : yes
cpuid level : 13
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 fma cx16 xtpr pdcm pcid sse4_1 sse4_2 movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm ida arat epb xsaveopt pln pts dtherm tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid
bogomips : 7000.78
clflush size : 64
cache_alignment : 64
address sizes : 39 bits physical, 48 bits virtual
building from/to local media. Any special setup I should use here for my test?
I build on a SSD-like disk (flash media on PCI Express slot), so I
should have almost no i/o time.
I build with 'make -j16'
If you have time having the output of turbostat for a build with and
without would be very useful.
$ ./turbostat
turbostat: no /dev/cpu/0/msr
Try "# modprobe msr": No such file or directory
Bah, I'll rebuild with msr and get you that info...
Attached is two files, "fast_build.txt" is when your patch is reverted,
and building a kernel takes only 2 minutes. "slow_build.txt" is when
the patch is not reverted, using Linus's latest tree, and I stopped it
after a while because I got bored :)
I think the numbers in these two files shows the real problem that is
happening...
For some reason intel_pstate is deciding you have no load and no increasing
the P state :-(
If there's anything else you need me to test, please let me know.
I are running on Linus's tree could you run the following script with the patch
applied and send the output of "perf script".
#! /bin/sh
sudo perf record -C 5 -c 1 -e power:pstate_sample&
sleep .5
foo=$!
taskset -c 5 cat /dev/zero > /dev/null&
bar=$!
sleep 1
kill $bar
sleep .5
sudo kill $foo
thanks,
greg k-h
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/