Re: Question about hugepages, shared memory, and /dev/shm....

Thanks for replies, all.

Should have mentioned, all environments are x86-64.  I've been reading lots of 
MOS docs, don't see anything to explain this.

To be clear, on the systems where I can see the files in /dev/shm, I have 
confirmed hugepages are in use via /proc/meminfo and looking for Large Pages 
message in alert log at instance startup.  I see hugepages *are* in use, and I 
*still* see files in /dev/shm, with timestamps that are consistent with (same 
or later than) instance uptime.

No idea what's going on here....

It's worth noting, though it may not be relevant, that I have (thus far) only 
observed this on our various eBS environments.

-Mark

From: Chris Taylor 
<christopherdtaylor1994@xxxxxxxxx<mailto:christopherdtaylor1994@xxxxxxxxx>>
Date: Wednesday, April 16, 2014 at 11:59 AM
To: Mark Bobak <Mark.Bobak@xxxxxxxxxxxx<mailto:Mark.Bobak@xxxxxxxxxxxx>>
Cc: "oracle-l@xxxxxxxxxxxxx<mailto:oracle-l@xxxxxxxxxxxxx>" 
<oracle-l@xxxxxxxxxxxxx<mailto:oracle-l@xxxxxxxxxxxxx>>
Subject: Re: Question about hugepages, shared memory, and /dev/shm....

Is there a difference between the architectures of the machines that have 
/dev/shm allotments and the machines that do not?  I see some references in the 
Oracle Docs about /dev/shm on x86 boxes so I was curious if 32-bit behaves 
differently than 64-bit boxes in this regard?

Chris


On Wed, Apr 16, 2014 at 10:47 AM, Mark Bobak 
<Mark.Bobak@xxxxxxxxxxxx<mailto:Mark.Bobak@xxxxxxxxxxxx>> wrote:
Hi All,

So, I thought I really understood this stuff, but I'm a little baffled here, 
and I wonder if anyone can offer me a clue?

Here's what I (think I) know:
1.)  AMM (setting memory_target) is *not* compatible with a hugepages 
configuration.  Any attempt to use hugepages will lock out the memory allocated 
to hugepages and AMM will only use non-hugepage memory allocations, the effect 
of which would be like removing the huge page allocated memory from the system.
2.)  ASMM (setting sga_target and pga_aggregate_target) and MMM (manually 
setting db_cache_size and pool sizes) *are* compatible with a hugepages 
configuration, and for any non-trivially sized SGA, hugepages is strongly 
recommended.
3.)  If hugepages are *not* configured, and AMM is used, memory segments will 
be mapped in /dev/shm.
4.)  If hugepages *are* used, no memory segments will be visible in /dev/shm.

So, that's what I think is true about memory configuration and hugepages 
configuration.

That seems to be consistent throughout our environment, which mostly has ASMM 
or MMM and hugepages configuration.

However, and this is where my confusion comes in, we have several eBS 
environments, which seem to have a valid and active hugepages configuration, 
are using ASMM (not AMM), and *still* I can see memory segments allocated in 
/dev/shm??  Any idea how this is possible?

Here's an example from our preprod environment:
(Content was too long for Oracle-L, so here's a paste bin URL)

http://pastebin.com/7w2V2jEa


So, I'm a little baffled here.  I thought these were mutually exclusive 
features.


Note also that the timestamps on the /dev/shm segments is *after* instance 
startup time, so, I don't think these are "orphan" memory segments....


Anyone out there can clue me in?


Thanks,


-Mark

Other related posts: