Re: [RFC PATCH 0/6] module, kbuild: Faster boot with custom kernel.

[Kay Sievers]

On Thu, Feb 19, 2009 at 21:48, Kay Sievers <kay.sievers@xxxxxxxx> wrote:
> On Thu, Feb 19, 2009 at 12:41, Kay Sievers <kay.sievers@xxxxxxxx> wrote:
>> On Thu, Feb 19, 2009 at 12:15, Rusty Russell <rusty@xxxxxxxxxxxxxxx> wrote:
>>> On Thursday 19 February 2009 00:27:58 Kay Sievers wrote:
>>>> some modules wait for 200-500 milliseconds to
>>>> get the lock released, some larger modules spend 50 milliseconds in
>>>> load_module(), many of them around 20 milliseconds.
>>>
>>> OK, this is an untested hack (don't try unloading modules, not sure symbol
>>> resolution isn't racy now I've killed the lock).  Does it change the numbers?
>>
>> That changes it dramatically. The numbers from the sycall until the
>> linked-in module are now down to 15-25 milliseconds, and for a few
>> large modules 50-100.
>>
>> (One crazy exception is ipv6, which takes 620 milliseconds to link, no
>> idea what it needs to do.)
>
> Sorry, this was caused by I/O wait from disk for that huge module, and
> gets to reasonable numbers by putting all modules into RAM before
> loading them.
>
>> I'll compare a few bootup times with and without the patch, and come
>> back later today with the real numbers.
>
> The whole massive parallel modprobe happens during udev coldplug. I
> tried a 2GHz Dual Core laptop, and a setup without initramfs here,
> which loads ~40 modules. All the kernel modules are copied to a ramfs
> mount before the coldplug is started.
>
> I measured the time from the first modprobe that happened in the
> kernel to the loading of "dummy", which I manually trigger from the
> udev boot script, and which gets called right after udev has settled
> and handled all events.
>
> With the mutex it takes 1.8 seconds, without it, it takes 1.3 seconds.
> If I comment out the creation of the stop_machine() threads, it gets
> down to 1.1 seconds.
>
> With the mutex, I see code waiting for up to 180 milliseconds waiting
> for the mutex, the average between 20-40 milliseconds.
>
> Without the mutex the largest time to link is 30 milliseconds, and
> most of them are around 5-10 milliseconds.
>
> Without the mutex and the stop_machine() creation, the flow of tracing
> output looks like real work and, depending on the actual module, they
> spend time in various stages of linking, relocation and so on, there
> are no long delays for any of the modules, like I see with the current
> code.
>
> It would be great, if we can safely minimize the time spent in the mutex.

Further testing revealed, if I only comment out the stop_machine()
preparation, which is used in an error case, I get almost the same
improvement, even with the original mutex in place. Without the mutex
it's still a bit better, maybe it would be much better if we have more
CPUs, but all the long delays are gone only with removing the
stop_machine() preparation.

Thanks,
Kay
--
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/