Re: Hugepages - benefits / drawbacks

From: Roman Podshivalov <roman.podshivalov_at_gmail.com>
Date: Wed, 8 Oct 2008 13:45:56 -0400
Message-ID: <55f303590810081045j43587250ya60937f239a16ca6@mail.gmail.com>


Hmmm,

Let me take that back, I guess I was too tired yesterday. I did same test today and indeed observed some mmap calls, but from my understanding, confirmed with pmap output what they do just space allocation in anon area and zeroing it by mapping to /dev/zero. fd #8 is /dev/zero in pfiles output from background process. Here there are snippets from truss:

15.9666 0.0007 mmap(0x00000000, 1048576, PROT_NONE, MAP_PRIVATE|MAP_NORESERVE, 8, 2031616) = 0xFFFFFFFF7ABF0000 15.9667 0.0001 mmap(0xFFFFFFFF7ABF0000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7ABF0000 15.9674 0.0007 mmap(0xFFFFFFFF7AC10000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7AC10000 15.9680 0.0006 mmap(0xFFFFFFFF7AC30000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7AC30000 15.9687 0.0007 mmap(0xFFFFFFFF7AC50000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7AC50000 15.9694 0.0007 mmap(0xFFFFFFFF7AC70000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7AC70000 15.9707 0.0013 mmap(0xFFFFFFFF7AC90000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7AC90000 15.9714 0.0007 mmap(0xFFFFFFFF7ACB0000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7ACB0000 15.9720 0.0006 mmap(0xFFFFFFFF7ACD0000, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 8, 0) = 0xFFFFFFFF7ACD0000

correspoiding pmap:

FFFFFFFF7ABF0000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7AC10000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7AC30000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7AC50000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7AC70000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7AC90000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7ACB0000 128 128 128 - 8K rw--- [ anon ]
FFFFFFFF7ACD0000 128 128 128 - 8K rw--- [ anon ]
thanks
--romas
On Tue, Oct 7, 2008 at 4:30 PM, Roman Podshivalov < roman.podshivalov_at_gmail.com> wrote:

> With default setting of _use_ism_for_pga set to TRUE - no changes in
> shared memory segments (pmap -s/pmap -x) were observed during sorting
> operation. Speaking of mmap/brk calls - no such calls happened _during the
> sort_ according to truss output. Looks like PGA is not using ISM despite the
> underscore parameter 8-). Oracle 10.2.0.3, Solaris 10.
>
> --romas
>
> On Tue, Oct 7, 2008 at 3:08 PM, Yong Huang <yong321_at_yahoo.com> wrote:
>
>> ipcs probably won't show anything interesting. Can you check truss output
>> when you do a big sort (select count(*) from dba_source order by 1)? Filter
>> on calls such as brk, mmap (get args of this). Also pmap -x | grep shmid.
>> See if the mapped size increases when you do big sorting. And, pmap -s <your
>> shadow process> to see which page size segment increases its size.
>>
>> Yong Huang
>>
>>
>> --- On Tue, 10/7/08, Roman Podshivalov <roman.podshivalov_at_gmail.com>
>> wrote:
>>
>> > From: Roman Podshivalov <roman.podshivalov_at_gmail.com>
>> > Subject: Re: Hugepages - benefits / drawbacks
>> > To: yong321_at_yahoo.com
>> > Cc: "Christo Kutrovsky" <kutrovsky.oracle_at_gmail.com>,
>> oracle-l_at_freelists.org, krishmanoh_at_gmail.com
>> > Date: Tuesday, October 7, 2008, 1:59 PM
>> >
>> > Yep that parameter caught my attention too. I just did a
>> > quick test and
>> > started database with this parameter set to true and false.
>> > No difference in
>> > ipcs -ma output was found....Could that be related to MTS
>> > configuration ?
>> >
>> > --romas
>> >
>> > On Tue, Oct 7, 2008 at 2:55 PM, Yong Huang
>> > <yong321_at_yahoo.com> wrote:
>> >
>> > > > On Tue, Oct 7, 2008 at 1:26 PM, Yong Huang
>> > > > <yong321_at_yahoo.com> wrote:
>> > > > > Hi Krishna,
>> > > > >
>> > > > >> on Solaris PGA uses ISM
>> > > > >
>> > > > > Wouldn't that cause PGA to be shared?
>> > >
>> > > Tanel Poder's presentation "Memory Management
>> > and Latching Improvements in
>> > > Oracle9i and 10g" mentions the parameter
>> > _use_ism_for_pga. It didn't
>> > > register in my mind when I read it some time ago. I
>> > don't have access to
>> > > Oracle on Solaris right now. Why does Oracle do this?
>> > It would be
>> > > interesting to see the effect.
>> > >
>> > > Yong Huang
>>
>>
>>
>>
>

--
http://www.freelists.org/webpage/oracle-l
Received on Wed Oct 08 2008 - 12:45:56 CDT

Original text of this message