lkml.org 
[lkml]   [2012]   [Jun]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: AutoNUMA15
On 06/21/2012 10:55 PM, Andrea Arcangeli wrote:

> On Thu, Jun 21, 2012 at 03:29:52PM +0800, Alex Shi wrote:
>>> I released an AutoNUMA15 branch that includes all pending fixes:
>>>
>>> git clone --reference linux -b autonuma15 git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git
>>>
>>
>> I did a quick testing on our
>> specjbb2005/oltp/hackbench/tbench/netperf-loop/fio/ffsb on NHM EP/EX,
>> Core2 EP, Romely EP machine, In generally no clear performance change
>> found. Is this results expected for this patch set?
>
> hackbench and network benchs won't get benefit (the former
> overschedule like crazy so there's no way any autonuma balancing can
> have effect with such an overscheduling and zillion of threads, the
> latter is I/O dominated usually taking so little RAM it doesn't
> matter, the memory accesses on the kernel side and DMA issue should
> dominate it in CPU utilization). Similar issue for filesystem
> benchmarks like fio.
>
> On all _system_ time dominated kernel benchmarks it is expected not to
> measure a performance optimization and if you don't measure a
> regression it's more than enough.
>
> The only benchmarks that gets benefit are userland where the user/nice
> time in top dominates. AutoNUMA cannot optimize or move kernel memory
> around, it only optimizes userland computations.
>
> So you should run HPC jobs. The only strange thing here is that
> specjbb2005 gets a measurable significant boost with AutoNUMA so if
> you didn't even get a boost with that you may want to verify:
>
> cat /sys/kernel/mm/autonuma/enabled == 1
>
> Also verify:
>
> CONFIG_AUTONUMA_DEFAULT_ENABLED=y
>
> If that's 1 well maybe the memory interconnect is so fast that there's
> no benefit?
>
> My numa01/02 benchmarks measures the best worst case of the hardware
> (not software), with -DINVERSE_BIND -DHARD_BIND parameters, you can
> consider running that to verify.


Could you like to give a url for the benchmarks?

>
> Probably there should be a little boot time kernel benchmark to
> measure the inverse bind vs hard bind performance across the first two
> nodes, if the difference is nil AutoNUMA should disengage and not even
> allocate the page_autonuma (now only 12 bytes per page but anyway).
>
> If you can retest with autonuma17 it would help too as there was some
> performance issue fixed and it'd stress the new autonuma migration lru
> code:
>
> git clone --reference linux -b autonuma17 git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git autonuma17
>
> And the very latest is always at the autonuma branch:
>
> git clone --reference linux -b autonuma git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git autonuma


I got the commit till 2c7535e100805d. and retested specjbb2005 with
jrockit and openjdk again on my Romely EP(2P * 8 cores * HT, with 64GB
memory). find the openjdk has about 2% regression, while jrockit has no
clear change.


the testing user 2 instances, each of them are pinned to a node. some
setting is here:
per_jvm_warehouse_rampup = 3.0
per_jvm_warehouse_rampdown = 20.0
jvm_instances = 2
deterministic_random_seed = false
ramp_up_seconds = 30
measurement_seconds = 240
starting_number_warehouses = 1
increment_number_warehouses = 1
ending_number_warehouses = 34
expected_peak_warehouse = 16

openjdk
java options:
-Xmx8g -Xms8g -Xincgc

jrockit use hugetlb and its options:
-Xmx8g -Xms8g -Xns4g -XXaggressive -Xlargepages -XXlazyUnlocking
-Xgc:genpar -XXtlasize:min=16k,preferred=64k

>
> Thanks,
> Andrea




\
 
 \ /
  Last update: 2012-06-26 10:42    [W:0.402 / U:0.024 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site