Huge pages on PPC64

[Gustavo Romero]

Hi Goetz,

On 25-10-2016 04:58, Lindenmaier, Goetz wrote:
> Hi Gustavo, 
> 
> At some point this was working, but we don't test this as 
> you always have to configure a machine to support large pages.
> The default configuration of our machines does not support it, see also 
> more /proc/meminfo | grep HugePages.
> So it might well be that it's no more working. 
> 
> Anyways, there is now transparent huge pages which works differently.
> Further, usually we run with 64k pages on ppc so that the page overhead is not
> as high as on x86 where you often have 4k pages.

I did some raw tests with SPECjbb2005 on my host just as a simple comparison. I
didn't take care of the optimal Hugepagesize, for instance, so absolute values
are not relevant.

$ uname -a
Linux pub 4.5.0-rc4-powerpc64le #1 SMP Debian 4.5~rc4-1~exp1 (2016-02-18) ppc64le GNU/Linux
$ free -m
              total        used        free      shared  buff/cache   available
Mem:         130579       77877       45850          19        6851       52502
Swap:         31279           0       31279
$ fgrep Huge /proc/meminfo
AnonHugePages:   8536064 kB
HugePages_Total:    4096
HugePages_Free:     2048
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:      16384 kB

Warnings on +UseHugeTLBFS were observed as I'm discussing with Volker and Thomas,
but no warnings popped up with +UseTransparentHugePages or +UseSHM even with
CompressedOops and CompressedClassPointers enabled.

No_HP:
java -Xms4096m -Xmx4096m spec.jbb.JBBmain -propfile SPECjbb.props

HugeTLBFS:
java -XX:+UseLargePages -Xms4096m -Xmx4096m spec.jbb.JBBmain -propfile SPECjbb.props

SHM:
java -XX:+UseLargePages -XX:+UseSHM -Xms4096m -Xmx4096m spec.jbb.JBBmain -propfile SPECjbb.props

THP:
java -XX:+UseLargePages -XX:+UseTransparentHugePages -Xms4096m -Xmx4096m spec.jbb.JBBmain -propfile SPECjbb.props


      No_HP     HugeTLBFS  SHM       THP
      --------- ---------  --------- ---------
run 1 453817    441585     458999    454443
run 2 445482    435336     465042    460181
run 3 454363    436282     459126    450166
run 4 454792    440744     458203    462308
run 5 449470    434870     458243    461129
mean  451584.8  437763.4   459922.6  457645.4
sd    4021.933  3160.189   2892.834  5159.975

Values are SPECjbb2005 scores after a valid run. I did 5 complete runs.


> So feel free to improve the situation.

I believe we can go with THP and SHM allocations, since HugeTLBFS was not that
impressive...

Thanks for your comments!


Regards,
Gustavo