Re: How extensive OpenBSD's write caching (for softdep or async-mounted UFS, as long as I never fsync() )?

[Otto Moerbeek]

On Sun, Feb 14, 2016 at 06:56:16AM -0500, Donald Allen wrote:

> On Sun, Feb 14, 2016 at 1:43 AM, Tinker  wrote:
> 
> > Did two tests, one with async and one with softdep, on amd64, 5.9-CURRENT,
> > UFS.
> >
> > (Checked "dd"'s sources and there is no fsync() anywhere in there.
> >
> > The bufcache setting was 90, 3GB free RAM, pushed 2GB of data using "dd"
> > to disk.
> >
> 
> Based on knowledge of Unices from long ago (not on direct knowledge of
> OpenBSD internals), I believe 'dd' uses raw I/O, which bypasses the buffer
> cache.
> 
> Those who know the details better than I should comment and correct me if
> I'm wrong, but if I'm right, your test doesn't prove anything about
> file-system write caching because your writes didn't go through the
> file-system.

Depends where you are writing to: a file or a device?

-Otto

> 
> /Don Allen
> 
> >
> > It took 12 and 15 seconds respectively, which is the harddrive's write
> > speed - the buffer cache of course would have absorbed this in 0 seconds.)
> >
> >
> > So, both runs showed that OpenBSD *not* does any write caching to talk
> > about, at all.
> >
> >
> > This means if a program wants write caching, it needs to implement it
> > itself.
> >
> > Good to know.
> >
> > Tinker
> >
> >
> > On 2016-02-13 23:47, Tinker wrote:
> >
> >> Hi,
> >>
> >> How much of my file writing, and filesystem operations such as
> >> creating a new file/directory, will land in OpenBSD's disk/write cache
> >> without touching the disk before return of the respective operation to
> >> my program, for softdep or async UFS media and I never fsync() ?
> >>
> >>
> >> This is relevant to know for any usecase where there may be a big
> >> write load to a magnet disk *and* there's lots of RAM and "sysctl
> >> kern.bufcachepercent" is high.
> >>
> >> If those ops will be done in a way that is synchronous with the magnet
> >> disk, the actual fopen(), fwrite(), fread() (for re-read of the data
> >> that's been written but still only is in the OS RAM CACHE) etc. might
> >> be so slow that a program would need to implement its own write cache
> >> for supporting even small spikes in write activity.
> >>
> >> Sorry for the fuss but neither "man" nor googling taught me anything.
> >>
> >> Thanks!!
> >> Tinker

Re: How extensive OpenBSD's write caching (for softdep or async-mounted UFS, as long as I never fsync() )?

[Tinker]

On 2016-02-14 19:20, Otto Moerbeek wrote:

On Sun, Feb 14, 2016 at 06:56:16AM -0500, Donald Allen wrote:


On Sun, Feb 14, 2016 at 1:43 AM, Tinker  wrote:

> Did two tests, one with async and one with softdep, on amd64, 5.9-CURRENT,
> UFS.
>
> (Checked "dd"'s sources and there is no fsync() anywhere in there.
>
> The bufcache setting was 90, 3GB free RAM, pushed 2GB of data using "dd"
> to disk.
>

Based on knowledge of Unices from long ago (not on direct knowledge of
OpenBSD internals), I believe 'dd' uses raw I/O, which bypasses the 
buffer

cache.

Those who know the details better than I should comment and correct me 
if

I'm wrong, but if I'm right, your test doesn't prove anything about
file-system write caching because your writes didn't go through the
file-system.


Depends where you are writing to: a file or a device?

-Otto



I wrote to a file, as I pointed out it was on UFS which was mounted 
either softdep or async.


And, this is where essentially no write caching was experienced, making 
me understand that OpenBSD essentially not does write caching (and the 
"buffer cache" technically is a read cache).

Re: How extensive OpenBSD's write caching (for softdep or async-mounted UFS, as long as I never fsync() )?

[Donald Allen]

On Sun, Feb 14, 2016 at 1:43 AM, Tinker  wrote:

> Did two tests, one with async and one with softdep, on amd64, 5.9-CURRENT,
> UFS.
>
> (Checked "dd"'s sources and there is no fsync() anywhere in there.
>
> The bufcache setting was 90, 3GB free RAM, pushed 2GB of data using "dd"
> to disk.
>

Based on knowledge of Unices from long ago (not on direct knowledge of
OpenBSD internals), I believe 'dd' uses raw I/O, which bypasses the buffer
cache.

Those who know the details better than I should comment and correct me if
I'm wrong, but if I'm right, your test doesn't prove anything about
file-system write caching because your writes didn't go through the
file-system.

/Don Allen

>
> It took 12 and 15 seconds respectively, which is the harddrive's write
> speed - the buffer cache of course would have absorbed this in 0 seconds.)
>
>
> So, both runs showed that OpenBSD *not* does any write caching to talk
> about, at all.
>
>
> This means if a program wants write caching, it needs to implement it
> itself.
>
> Good to know.
>
> Tinker
>
>
> On 2016-02-13 23:47, Tinker wrote:
>
>> Hi,
>>
>> How much of my file writing, and filesystem operations such as
>> creating a new file/directory, will land in OpenBSD's disk/write cache
>> without touching the disk before return of the respective operation to
>> my program, for softdep or async UFS media and I never fsync() ?
>>
>>
>> This is relevant to know for any usecase where there may be a big
>> write load to a magnet disk *and* there's lots of RAM and "sysctl
>> kern.bufcachepercent" is high.
>>
>> If those ops will be done in a way that is synchronous with the magnet
>> disk, the actual fopen(), fwrite(), fread() (for re-read of the data
>> that's been written but still only is in the OS RAM CACHE) etc. might
>> be so slow that a program would need to implement its own write cache
>> for supporting even small spikes in write activity.
>>
>> Sorry for the fuss but neither "man" nor googling taught me anything.
>>
>> Thanks!!
>> Tinker

Re: How extensive OpenBSD's write caching (for softdep or async-mounted UFS, as long as I never fsync() )?

[Tinker]

Did two tests, one with async and one with softdep, on amd64, 
5.9-CURRENT, UFS.


(Checked "dd"'s sources and there is no fsync() anywhere in there.

The bufcache setting was 90, 3GB free RAM, pushed 2GB of data using "dd" 
to disk.


It took 12 and 15 seconds respectively, which is the harddrive's write 
speed - the buffer cache of course would have absorbed this in 0 
seconds.)



So, both runs showed that OpenBSD *not* does any write caching to talk 
about, at all.



This means if a program wants write caching, it needs to implement it 
itself.


Good to know.

Tinker

On 2016-02-13 23:47, Tinker wrote:

Hi,

How much of my file writing, and filesystem operations such as
creating a new file/directory, will land in OpenBSD's disk/write cache
without touching the disk before return of the respective operation to
my program, for softdep or async UFS media and I never fsync() ?


This is relevant to know for any usecase where there may be a big
write load to a magnet disk *and* there's lots of RAM and "sysctl
kern.bufcachepercent" is high.

If those ops will be done in a way that is synchronous with the magnet
disk, the actual fopen(), fwrite(), fread() (for re-read of the data
that's been written but still only is in the OS RAM CACHE) etc. might
be so slow that a program would need to implement its own write cache
for supporting even small spikes in write activity.

Sorry for the fuss but neither "man" nor googling taught me anything.

Thanks!!
Tinker