Обсуждение: Re: VX_CONCURRENT flag on vxfs( 5.1 or later) for performance for postgresql?

On 04/30/2011 12:24 AM, Hsien-Wen Chu wrote:
> I'm little bit confuse why it is not safe. and my question is following.
>
> for database application, we need to avoid double cache, PostgreSQL
> shared_buffer will cache the data, so we do not want to file system to
> cache the data right?. so  the DIRECT IO is better, right?.
>

No.  There are parts of PostgreSQL that expect the operating system to
do write caching.  Two examples are the transaction logs and the
processing done by VACUUM.  If you eliminate that with direct I/O, the
slowdown can be much, much larger than what you gain by eliminating
double-buffering on reads.

On the read side, PostgreSQL also expects that operating system features
like read-ahead are working properly.  While this does introduce some
double-buffering, the benefits for sequential scans are larger than that
overhead, too.  You may not get the expected read-ahead behavior if you
use direct I/O.

Direct I/O is not a magic switch that makes things faster; you have to
very specifically write your application to work around what it does,
good and bad, before it is expected to improves things.  And PostgreSQL
isn't written that way.  It definitely requires OS caching to work well.

> for VXFS, if the we use ioctl(fd,vx_cacheset,vx_concurrent) API,
> according to the vxfs document, it will hold a shared lock for write
> operation, but not the exclusive clock, also it is a direct IO,
>

There are very specific technical requirements that you must follow when
using direct I/O.  You don't get direct I/O without also following its
alignment needs.  Read the "Direct I/O best practices" section of
http://people.redhat.com/msnitzer/docs/io-limits.txt for a quick intro
to the subject.  And there's this additional set of requirements you
mention in order for this particular VXFS feature to work, which I can't
even comment on.  But you can be sure PostgreSQL doesn't try to do
either of those things--it's definitely not aligning for direct I/O.
Has nothing to do with ACID or the filesystem.

Now, the VXFS implementation may do some tricks that bypass the
alignment requirements.  But even if you got it to work, it would still
be slower for anything but some read-only workloads.  Double buffering
is really not that big of a performance problem, you just need to make
sure you don't set shared_buffers to an extremely large value.

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

Hi Mr. Greg Smith

since the block size is 8k for the default, and it consisted with many
tuple/line; as my understand, if any tuple/line is changed(maybe
update, insert, delete). the block will be marked as dirty block. and
then it will be flashed to disk by bgwriter.

so my question is if the data block(8k) is aligned with the file
system block?  if it is aligned with file system block, so what's the
potential issue make it is not safe for direct io. (please  assume
vxfs, vxvm and the disk sector is aligned ).please correct me if any
incorrect.

thank you very much
Tony


On 4/30/11, Greg Smith <greg@2ndquadrant.com> wrote:
> On 04/30/2011 12:24 AM, Hsien-Wen Chu wrote:
>> I'm little bit confuse why it is not safe. and my question is following.
>>
>> for database application, we need to avoid double cache, PostgreSQL
>> shared_buffer will cache the data, so we do not want to file system to
>> cache the data right?. so  the DIRECT IO is better, right?.
>>
>
> No.  There are parts of PostgreSQL that expect the operating system to
> do write caching.  Two examples are the transaction logs and the
> processing done by VACUUM.  If you eliminate that with direct I/O, the
> slowdown can be much, much larger than what you gain by eliminating
> double-buffering on reads.
>
> On the read side, PostgreSQL also expects that operating system features
> like read-ahead are working properly.  While this does introduce some
> double-buffering, the benefits for sequential scans are larger than that
> overhead, too.  You may not get the expected read-ahead behavior if you
> use direct I/O.
>
> Direct I/O is not a magic switch that makes things faster; you have to
> very specifically write your application to work around what it does,
> good and bad, before it is expected to improves things.  And PostgreSQL
> isn't written that way.  It definitely requires OS caching to work well.
>
>> for VXFS, if the we use ioctl(fd,vx_cacheset,vx_concurrent) API,
>> according to the vxfs document, it will hold a shared lock for write
>> operation, but not the exclusive clock, also it is a direct IO,
>>
>
> There are very specific technical requirements that you must follow when
> using direct I/O.  You don't get direct I/O without also following its
> alignment needs.  Read the "Direct I/O best practices" section of
> http://people.redhat.com/msnitzer/docs/io-limits.txt for a quick intro
> to the subject.  And there's this additional set of requirements you
> mention in order for this particular VXFS feature to work, which I can't
> even comment on.  But you can be sure PostgreSQL doesn't try to do
> either of those things--it's definitely not aligning for direct I/O.
> Has nothing to do with ACID or the filesystem.
>
> Now, the VXFS implementation may do some tricks that bypass the
> alignment requirements.  But even if you got it to work, it would still
> be slower for anything but some read-only workloads.  Double buffering
> is really not that big of a performance problem, you just need to make
> sure you don't set shared_buffers to an extremely large value.
>
> --
> Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
> PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
> "PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books
>
>

On Sat, Apr 30, 2011 at 4:51 AM, Hsien-Wen Chu <chu.hsien.wen@gmail.com> wrote:
> since the block size is 8k for the default, and it consisted with many
> tuple/line; as my understand, if any tuple/line is changed(maybe
> update, insert, delete). the block will be marked as dirty block. and
> then it will be flashed to disk by bgwriter.

True...

> so my question is if the data block(8k) is aligned with the file
> system block?  if it is aligned with file system block, so what's the
> potential issue make it is not safe for direct io. (please  assume
> vxfs, vxvm and the disk sector is aligned ).please correct me if any
> incorrect.

It's not about safety - it's about performance.  On a machine with
64GB of RAM, a typical setting for shared_buffers set to 8GB.  If you
start reading blocks into the PostgreSQL cache - or writing them out
of the cache - in a way that bypasses the filesystem cache, you're
going to have only 8GB of cache, instead of some much larger amount.
More cache = better performance.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

You should rather consider VxFS tuning - it has an auto-discovery for
DIRECT I/O according the block size. Just change this setting to 8K or
16-32K depending on your workload - then all I/O operations with a
bigger block size will be executed in DIRECT mode and bypass FS cache
(which logical as usually it'll correspond to a full scan or a seq
scan of some data), while I/O requests with smaller blocks will remain
cached which is very useful as it'll mainly cache random I/O (mainly
index access)..

With such a tuning I've got over %35 performance improvement comparing
to any other states (full DIRECT or fully cached).

Rgds,
-Dimitri


Rgds,
-Dimitri

On 5/5/11, Robert Haas <robertmhaas@gmail.com> wrote:
> On Sat, Apr 30, 2011 at 4:51 AM, Hsien-Wen Chu <chu.hsien.wen@gmail.com>
> wrote:
>> since the block size is 8k for the default, and it consisted with many
>> tuple/line; as my understand, if any tuple/line is changed(maybe
>> update, insert, delete). the block will be marked as dirty block. and
>> then it will be flashed to disk by bgwriter.
>
> True...
>
>> so my question is if the data block(8k) is aligned with the file
>> system block?  if it is aligned with file system block, so what's the
>> potential issue make it is not safe for direct io. (please  assume
>> vxfs, vxvm and the disk sector is aligned ).please correct me if any
>> incorrect.
>
> It's not about safety - it's about performance.  On a machine with
> 64GB of RAM, a typical setting for shared_buffers set to 8GB.  If you
> start reading blocks into the PostgreSQL cache - or writing them out
> of the cache - in a way that bypasses the filesystem cache, you're
> going to have only 8GB of cache, instead of some much larger amount.
> More cache = better performance.
>
> --
> Robert Haas
> EnterpriseDB: http://www.enterprisedb.com
> The Enterprise PostgreSQL Company
>
> --
> Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
> To make changes to your subscription:
> http://www.postgresql.org/mailpref/pgsql-performance
>