orangefs: complete Christoph's "remember count" reversion.
diff mbox series

Message ID 20200404162826.181808-1-hubcap@kernel.org
State New
Headers show
Series
  • orangefs: complete Christoph's "remember count" reversion.
Related show

Commit Message

hubcap@kernel.org April 4, 2020, 4:28 p.m. UTC
From: Mike Marshall <hubcap@omnibond.com>

---------- added text for this mail message ----------------
Here's a related patch. Something like this needs to be added
to Christoph's reversion in order for Orangefs to compile and
keep working. Using this static read size rather than using
the reverted logic to determine a read size seems faster in
my tests anyhow.

I accept Christoph's assertion that there would be a race, and
I looked at some of the vfs code (vfs_read, new_sync_read and
related). I guess the race Christoph sees would happen in
a threaded userspace program? I would be a better maintainer
if I saw the race more clearly, if anyone wants to go into
it. 

I guess I could employ a locking scheme if I wanted keep the
reverted code (I don't) instead of getting rid of it?

As an aside, the page cache has been a blessing and a curse
for us. Since we started using it, small IO has improved
incredibly, but our max speed hits a plateau before it otherwise
would have. I think because of all the page size copies we have
to do to fill our 4 meg native buffers. I try to read about all
the new work going into the page cache in lwn, and make some
sense of the new code :-). One thing I remember is when
Christoph Lameter said "the page cache does not scale", but
I know the new work is focused on that. If anyone has any
thoughts about how we could make improvments on filling our
native buffers from the page cache (larger page sizes?),
feel free to offer any help...

Anywho... thanks and I'll try to get this patch and
Christoph's two pulled if nobody sees a problem with them...

-Mike
------------------------------------------------------------

Logically, optimal Orangefs "pages" are 4 megabytes. Reading large Orangefs
files 4096 bytes at a time is like trying to kick a dead whale
down the beach. Before Christoph's "Revert orangefs: remember
count when reading." I tried to give users a knob whereby they could,
for example, use "count" in read(2) or bs with dd(1) to get
whatever they considered an appropriate amount of bytes at
a time from Orangefs and fill as many page cache pages as they
could at once.

Without the racy code that Christoph reverted Orangefs won't
even compile, much less work. So this replaces the logic that
used the private file data that Christoph reverted with
a static number of bytes to read from Orangefs.

I ran tests like the following to determine what a
reasonable static number of bytes might be:

dd if=/pvfsmnt/asdf of=/dev/null count=128 bs=4194304
dd if=/pvfsmnt/asdf of=/dev/null count=256 bs=2097152
dd if=/pvfsmnt/asdf of=/dev/null count=512 bs=1048576
                          .
                          .
                          .
dd if=/pvfsmnt/asdf of=/dev/null count=4194304 bs=128

Reads seem faster using the static number, so my "knob code"
wasn't just racy, it wasn't even a good idea...

Signed-off-by: Mike Marshall <hubcap@omnibond.com>
---
 fs/orangefs/inode.c | 39 ++++++---------------------------------
 1 file changed, 6 insertions(+), 33 deletions(-)

Comments

Matthew Wilcox April 4, 2020, 5:43 p.m. UTC | #1
On Sat, Apr 04, 2020 at 12:28:26PM -0400, hubcap@kernel.org wrote:
> As an aside, the page cache has been a blessing and a curse
> for us. Since we started using it, small IO has improved
> incredibly, but our max speed hits a plateau before it otherwise
> would have. I think because of all the page size copies we have
> to do to fill our 4 meg native buffers. I try to read about all
> the new work going into the page cache in lwn, and make some
> sense of the new code :-). One thing I remember is when
> Christoph Lameter said "the page cache does not scale", but
> I know the new work is focused on that. If anyone has any
> thoughts about how we could make improvments on filling our
> native buffers from the page cache (larger page sizes?),
> feel free to offer any help...

Umm, 4MB native buffers are ... really big ;-)  I wasn't planning on
going past PMD_SIZE (ie 2MB on x86) for the readahead large pages,
but if a filesystem wants that, then I should change that plan.

What I was planning for, but don't quite have an implementation nailed
down for yet, is allowing filesystems to grow the readahead beyond that
wanted by the generic code.  Filesystems which implement compression
frequently want blocks in the 256kB size range.  It seems like OrangeFS
would fit with that scheme, as long as I don't put a limit on what the
filesystem asks for.

So yes, I think within the next year, you should be able to tell the
page cache to allocate 4MB pages.  You will still need a fallback path
for when memory is too fragmented to allocate new pages, but if you're
using 4MB pages, then hopefully we'll be able to reclaim a clean 4MB
pages from elsewhere in the page cache and supply you with a new one.
Mike Marshall April 4, 2020, 8:57 p.m. UTC | #2
Matthew>  So yes, I think within the next year, you should be
Matthew>  able to tell the page cache to allocate 4MB pages.

I can't find the ascii thumbs up emoji :-) ...

-Mike

On Sat, Apr 4, 2020 at 1:43 PM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Sat, Apr 04, 2020 at 12:28:26PM -0400, hubcap@kernel.org wrote:
> > As an aside, the page cache has been a blessing and a curse
> > for us. Since we started using it, small IO has improved
> > incredibly, but our max speed hits a plateau before it otherwise
> > would have. I think because of all the page size copies we have
> > to do to fill our 4 meg native buffers. I try to read about all
> > the new work going into the page cache in lwn, and make some
> > sense of the new code :-). One thing I remember is when
> > Christoph Lameter said "the page cache does not scale", but
> > I know the new work is focused on that. If anyone has any
> > thoughts about how we could make improvments on filling our
> > native buffers from the page cache (larger page sizes?),
> > feel free to offer any help...
>
> Umm, 4MB native buffers are ... really big ;-)  I wasn't planning on
> going past PMD_SIZE (ie 2MB on x86) for the readahead large pages,
> but if a filesystem wants that, then I should change that plan.
>
> What I was planning for, but don't quite have an implementation nailed
> down for yet, is allowing filesystems to grow the readahead beyond that
> wanted by the generic code.  Filesystems which implement compression
> frequently want blocks in the 256kB size range.  It seems like OrangeFS
> would fit with that scheme, as long as I don't put a limit on what the
> filesystem asks for.
>
> So yes, I think within the next year, you should be able to tell the
> page cache to allocate 4MB pages.  You will still need a fallback path
> for when memory is too fragmented to allocate new pages, but if you're
> using 4MB pages, then hopefully we'll be able to reclaim a clean 4MB
> pages from elsewhere in the page cache and supply you with a new one.

Patch
diff mbox series

diff --git a/fs/orangefs/inode.c b/fs/orangefs/inode.c
index 961c0fd8675a..fb0884626d18 100644
--- a/fs/orangefs/inode.c
+++ b/fs/orangefs/inode.c
@@ -259,46 +259,19 @@  static int orangefs_readpage(struct file *file, struct page *page)
 	pgoff_t index; /* which page */
 	struct page *next_page;
 	char *kaddr;
-	struct orangefs_read_options *ro = file->private_data;
 	loff_t read_size;
-	loff_t roundedup;
 	int buffer_index = -1; /* orangefs shared memory slot */
 	int slot_index;   /* index into slot */
 	int remaining;
 
 	/*
-	 * If they set some miniscule size for "count" in read(2)
-	 * (for example) then let's try to read a page, or the whole file
-	 * if it is smaller than a page. Once "count" goes over a page
-	 * then lets round up to the highest page size multiple that is
-	 * less than or equal to "count" and do that much orangefs IO and
-	 * try to fill as many pages as we can from it.
-	 *
-	 * "count" should be represented in ro->blksiz.
-	 *
-	 * inode->i_size = file size.
+	 * Get up to this many bytes from Orangefs at a time and try
+	 * to fill them into the page cache at once.
+	 * Tests with dd made this seem like a reasonable static
+	 * number, if there was interest perhaps this number could
+	 * be made setable through sysfs...
 	 */
-	if (ro) {
-		if (ro->blksiz < PAGE_SIZE) {
-			if (inode->i_size < PAGE_SIZE)
-				read_size = inode->i_size;
-			else
-				read_size = PAGE_SIZE;
-		} else {
-			roundedup = ((PAGE_SIZE - 1) & ro->blksiz) ?
-				((ro->blksiz + PAGE_SIZE) & ~(PAGE_SIZE -1)) :
-				ro->blksiz;
-			if (roundedup > inode->i_size)
-				read_size = inode->i_size;
-			else
-				read_size = roundedup;
-
-		}
-	} else {
-		read_size = PAGE_SIZE;
-	}
-	if (!read_size)
-		read_size = PAGE_SIZE;
+	read_size = 524288;
 
 	if (PageDirty(page))
 		orangefs_launder_page(page);