Re: [Suspend-devel] Resume performance

[Anders Aagaard <aagaande@gmail.com>]

Rafael J. Wysocki wrote:
> On Tuesday, 9 of September 2008, Anders Aagaard wrote:
>> Rafael J. Wysocki wrote:
>>> On Sunday, 7 of September 2008, Anders Aagaard wrote:
>>>> Rafael J. Wysocki wrote:
>>>>> On Friday, 5 of September 2008, Anders Aagaard wrote:
>>>>>> Rafael J. Wysocki wrote:
>>>>>>> On Friday, 5 of September 2008, Anders Aagaard wrote:
>>>>>>>> Hi
>>>>>>> Hi,
>>>>>>>
>>>>>>> This is a kernel problem, so let's CC the LKML.
>>>>>>>
>>>>>>>> I have a intel P35 board with a quad core cpu in it, it's currently 
>>>>>>>> running as a server for a small network, and I'd like to be able to shut 
>>>>>>>> it down when idle, and use wake on lan to wake it up when it's needed. 
>>>>>>>> Now I got that part working quite well, but for some reason I have a 
>>>>>>>> long delay in resume.
>>>>>>>>
>>>>>>>> I seem to remember being able to resume this computer in 2-3 seconds 
>>>>>>>> when I was testing it, now it needs 35 seconds to resume.  It seems 
>>>>>>>> regardless of resume options used, and it always resumes to a working 
>>>>>>>> state without problems.
>>>>>>> What kernel are you using at the moment and which one was used for the
>>>>>>> testing?
>>>>>> I'm using gentoo's 2.6.25-r7, I've also tried vanilla sources.
>>>>> Would it be possible to test 2.6.27-rc5-gi7 from kernel.org?
>>>> Tested, makes no difference.
>>>>
>>>>>>>> I've tried quite a lot of things, booting with noapic/nosmp, booting a 
>>>>>>>> kernel without usb/network drivers, disabling ahci (using ata_piix 
>>>>>>>> driver instead of ahci), and there's always that one long delay.  And 
>>>>>>>> I'm not quite sure how the kernel printk timing information works, so 
>>>>>>>> I'm not sure whats causing that delay.
>>>>>>>>
>>>>>>>> Output from dmesg when booting with nosmp (to get accurate timing data):
>>>>>>>> scripts/show_delta -b "Force enabled HPET at resume"
>>>>>>>> [349.821150 < 7.039261 >] ata3.00: configured for UDMA/133
>>>>>>>> [349.821160 < 7.039271 >] sd 2:0:0:0: [sdc] 976773168 512-byte hardware 
>>>>>>>> sectors (500108 MB)
>>>>>>>> [349.821165 < 7.039276 >] sd 2:0:0:0: [sdc] Write Protect is off
>>>>>>>> [349.821166 < 7.039277 >] sd 2:0:0:0: [sdc] Mode Sense: 00 3a 00 00
>>>>>>>> [349.821173 < 7.039284 >] sd 2:0:0:0: [sdc] Write cache: enabled, read 
>>>>>>>> cache: enabled, doesn't support DPO or FUA
>>>>>>>> [349.972801 < 7.190912 >] ata1: SATA link up 3.0 Gbps (SStatus 123 
>>>>>>>> SControl 300)
>>>>>>>> [349.979060 < 7.197171 >] ata1.00: configured for UDMA/133
>>>>>>>> [349.979070 < 7.197181 >] sd 0:0:0:0: [sda] 976771055 512-byte hardware 
>>>>>>>> sectors (500107 MB)
>>>>>>>> [349.979075 < 7.197186 >] sd 0:0:0:0: [sda] Write Protect is off
>>>>>>>> [349.979076 < 7.197187 >] sd 0:0:0:0: [sda] Mode Sense: 00 3a 00 00
>>>>>>>> [349.979083 < 7.197194 >] sd 0:0:0:0: [sda] Write cache: enabled, read 
>>>>>>>> cache: enabled, doesn't support DPO or FUA
>>>>>>> It looks like this happens here.  Can you try to unload the network driver
>>>>>>> before suspend, please?
>>>>>> I tried to build a kernel without it, and it still takes the exact same 
>>>>>> amount to boot, I've also tried unloading usb drivers and it takes the 
>>>>>> exact same amount of time.
>>>>> Can you try to boot with init=/bin/bash and suspend to RAM?  (Please have a
>>>>> look at section 2 of Documentation/power/basic-pm-debugging.txt in the newer
>>>>> kernel sources).
>>>> I checked without X before, but forgot to unload the nvidia module, that 
>>>> apparently makes a big difference, I did some numbers with 
>>>> scripts/show_delta -b "Back to C".
>>>>
>>>> Nvidia and X : 32 seconds
>>>> No X (same result as booting with init=/bin/bash) : 8.3 seconds
>>>> Git kernel : 8.2 seconds
>>>> Light kernel (no sound, network card and usb drivers) : 8.17 seconds
>>>> ATI card instead of nvidia : 8.22 seconds
>>>>
>>>> I think we found the problem, I already replaced nvidia hardware in one 
>>>> computer to resolve another issue.  Really appreciate your help on this 
>>>> issue, this resume time works pretty well for me, it was a bit 
>>>> ridiculous when I could boot faster than resume.
>>>>
>>>> Is 8 seconds fairly expected?
>>> Well, it should be a bit less than that.  Usually, the resume shouldn't take
>>> more than 5 sec.
>>>
>>>> My other computer (same ati card) boots  
>>>> in about 2 seconds, but there's a lot less hardware in it (6 hd's and a 
>>>> ton of usb devices in one computer, 1 hd and 1 usb device in the other).
>>> That may matter a lot.  It would be interesting to check if detaching any of
>>> those devices from the first machine helps. ;-)
>>>
>>>>   I checked cold booting with and without usb devices, my light kernel 
>>>> boots to /bin/bash in 2.5 seconds, normal kernel about 7-8.  But I dont 
>>>> see anything about usb on resume.
>>> Of course the USB devices are also resumed and that takes time (comparable
>>> to the boot time).
>> Pulling out all my usb devices takes the resume time to 30.4
> 
> This is with the NVidia driver I guess?  So your resume appears to be 1.6 s
> faster without the USB devices.  Perhaps the disks also add to the latency.
> 
> If you have CONFIG_PCIEASPM set in the kernel config, you can try to unset it
> and see if that changes anything.

I did some more testing, including the old results for referance:
Nvidia and X : 32 seconds
No X (same result as booting with init=/bin/bash) : 8.3 seconds
Git kernel : 8.2 seconds
Light kernel (no sound, network card and usb drivers) : 8.17 seconds
ATI card instead of nvidia : 8.22 seconds
Nvidia - usb : 30.4

Latest tests:
ati-1hd : 5.96s
ati-1hd-nousb-noups : 5.92s
ati-1hd-bios disabled firewire,network,usb : 5.74s

I was sure it was the hd's that caused the latency, oh well, in either 
case it's a massive improvement from what I had.

Thanks a lot for your help, Anders.


> 
> Thanks,
> Rafael
>