RE: [PATCH] mm: moving dirty pages balancing to pdfludh entirely

["Ananiev, Leonid I" <leonid.i.ananiev@intel.com>]

Nikita Danilov writes:
I've used with " patched UP" to underline fast, simple, superficial
fixing.

> and small number of pdflush threads is able to achieve necessary IO
concurrency level.
Can set MAX_PDFLUSH_THREADS=1000 if it is as you say. You will see huge
number of created pdflush threads created. That is why I consider that
MAX_PDFLUSH_THREADS=8 is fast-fix patch-up which is used now as an
argument contra parallelism.

> If asynchronous write-out cannot cope with the rate of dirtying,
> synchronous write-out starts.
Quite the contrary in current design: synchronous write-out is started
first and after pdflush is waked up and it sees that all done.

> You, on the other hand, are trying to use pdflush for the goal it
wasn't
> designed for.
Just after patching pdflush is used REALLY for writing out of dirty
pages concurrently with user thread work.

> No wonder it doesn't work.

It doesn't work for you because you have not run it. Have you tried to
run any real kernel with proposed path and without? 

> You didn't look carefully enough. :-)
> ratelimit_pages = 32, hence, sync_writeback_pages() returns 48. It
could
> be different, of course, if ratelimit_pages were.

I ask one more: why is it 48. Does it make best performance for any
configuration? You explanation is: 48 because it is 32+16. Is it
explanation? Is it proving?

This patch makes performance benefits for current REAL hardware and
software state and opens new performance benefits in smp. The patch was
carefully benchmarked on several configurations: 1-8 cpu; 1-8Gb memory;
1-15 disks.
You have proposed tens far-fetched contra arguments. I have explained
carefully that each your argument is ungrounded.

Leonid
-----Original Message-----
From: Nikita Danilov [mailto:nikita@clusterfs.com] 
Sent: Thursday, July 06, 2006 9:37 PM
To: Ananiev, Leonid I
Cc: Bret Towe; Linux Kernel Mailing List
Subject: Re: [PATCH] mm: moving dirty pages balancing to pdfludh
entirely

Ananiev, Leonid I writes:
 > Nikita Danilov writes:
 > >> by introducing MAX_PDFLUSH_THREADS=8. Why just 8?
 > > Sorry, I don't understand. pdflush.c appeared in 2.5.8 and
 > Thanks for explanation. Why just 8? Answer: it was introduced in
2.5.8.

No, no, I was commenting on your (incorrect) statement that kernel "was
patched up by introducing MAX_PDFLUSH_THREADS=8". It never was: this
limit is part of the original design, not some "patch".

 > Why the constant MAX_PDFLUSH_THREADS is needed? Is it because current
 > kernel may create huge number of pdflush threads and overload the
 > system? Why we can not set MAX_PDFLUSH_THREADS=512? 1000?

Because this is not needed for current pdflush usage patterns. pdflush
is used for light background write-out, and small number of pdflush
threads is able to achieve necessary IO concurrency level, because they
are mostly non-blocking, thanks to the current_is_pdflush() checks. If
asynchronous write-out cannot cope with the rate of dirtying,
synchronous write-out starts.

You, on the other hand, are trying to use pdflush for the goal it wasn't
designed for. No wonder it doesn't work.

 > 
 > >> Why 48?
 > > This is explained in comment just above sync_writeback_pages()
 > > definition. Basically, ratelimit_pages pages might be dirtied
between
 > > calls to balance_dirty_pages(), and the latter tries to write out
 > *more*
 > > pages to keep number of dirty pages under control: negative
feedback
 > > control loop, of sorts.
 > 
 > I had asked "why 48?" is hard coded for any configuration. I do not
see
 > "48" in your explanation.

You didn't look carefully enough. :-)


/*
 * When balance_dirty_pages decides that the caller needs to perform
some
 * non-background writeback, this is how many pages it will attempt to
write.
 * It should be somewhat larger than RATELIMIT_PAGES to ensure that
reasonably
 * large amounts of I/O are submitted.
 */
static inline long sync_writeback_pages(void)
{
	return ratelimit_pages + ratelimit_pages / 2;
}

ratelimit_pages = 32, hence, sync_writeback_pages() returns 48. It could
be different, of course, if ratelimit_pages were.

 > 
 > You do not recommend to use hard coded constants but
 > MAX_PDFLUSH_THREADS=8 and write_chunk=48 are sacred according you
mind. 

No, I think the fewer hard-coded constants are here---the better. I'd
happily eliminate MAX_PDFLUSH_THREADS, but just dropping this constraint
is a non-solution in search of a problem, obviously.

But this discussion is degenerating, so I shall participate in it no
longer.

 > 
 > Leonid

Nikita.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/