[dpdk-dev] [PATCH] app/testpmd: adds mlockall() to fix pages
Aaron Conole
aconole at redhat.com
Wed Sep 13 14:28:42 CEST 2017
Thomas Monjalon <thomas at monjalon.net> writes:
> 12/09/2017 22:29, Aaron Conole:
>> Thomas Monjalon <thomas at monjalon.net> writes:
>>
>> > 12/09/2017 16:50, Aaron Conole:
>> >> Eelco Chaudron <echaudro at redhat.com> writes:
>> >>
>> >> > Call the mlockall() function, to attempt to lock all of its process
>> >> > memory into physical RAM, and preventing the kernel from paging any
>> >> > of its memory to disk.
>> >> >
>> >> > When using testpmd for performance testing, depending on the code path
>> >> > taken, we see a couple of page faults in a row. These faults effect
>> >> > the overall drop-rate of testpmd. On Linux the mlockall() call will
>> >> > prefault all the pages of testpmd (and the DPDK libraries if linked
>> >> > dynamically), even without LD_BIND_NOW.
>> >> >
>> >> > Signed-off-by: Eelco Chaudron <echaudro at redhat.com>
>> >>
>> >> Acked-by: Aaron Conole <aconole at redhat.com>
>> >
>> > It is interesting, but why make it in testpmd?
>> >
>> > Maybe it should be documented in this guide:
>> > http://dpdk.org/doc/guides/linux_gsg/nic_perf_intel_platform.html
>>
>> Well, I'm not sure what the user would be able to do to get the
>> prefaulting performance without having a library they use with
>> LD_PRELOAD and a function with the constructor attribute which does the
>> same thing, AND export LD_BIND_NOW before linking starts.
>>
>> The LD_BIND_NOW simply does the symbol resolution, but there's no
>> guarantee that it will fault all the code pages in to process space, and
>> without an mlockall(), I'm not sure that there's any kind of guarantee
>> that they don't get swapped out of resident memory (which also leads to
>> later page faults).
>>
>> Maybe I misunderstood the question?
>
> Maybe you misunderstood :)
>
> I was saying that if this improvement applies to applications,
> it should be documented in the tuning guide.
Ahh, okay. Yep, I agree with all you've written above.
More information about the dev
mailing list