lkml.org 
[lkml]   [2004]   [Sep]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patches in this message
/
Date
From
SubjectRe: [PATCH/RFC] Simplified Readahead
On Mon, 27 Sep 2004, Steven Pratt wrote:

> Ram Pai wrote:
>
> >On Mon, 2004-09-27 at 08:30, Steven Pratt wrote:
> >
> >
> >>Ram Pai wrote:
> >>
> >>
> >>>On Fri, 24 Sep 2004, Ram Pai wrote:
> >>>
> >>>
> >>>
> >>>>On Thu, 23 Sep 2004, Steven Pratt wrote:
> >>>>
> >>>>
> >>>>
> >
> > ..snip..
> >
> >
> >>>To summarize you noticed 3 problems:
> >>>
> >>>1. page cache hits not handled properly.
> >>>2. readahead thrashing not accounted.
> >>>3. read congestion not accounted.


I have enclosed 5 patches that address each of the issues.

1 . Code is obtuse and hard to maintain

The best I could do is update the comments to reflect the
current code. Hopefully that should help.

attached patch 1_comment.patch takes care of that part to
some extent.


2. page cache hits not handled properly.

I fixed this by decrementing the size of the next readahead window
by the number of pages hit in the page cache. Now it slowly
accomodates the page cache hits.

attached patch 2_cachehits.patch takes care of this issue.

3. queue congestion not handled.

The fix is: call force_page_cache_readahead() if we are
populating pages in the current window.
And call do_page_cache_readahead() if we are populating
pages in the ahead window. However if do_page_cache_readahead()
return with congestion, the readahead window is collapsed back
to size zero. This will ensure that the next time ahead window
is attempted to populate.

attached patch 3_queuecongestion.patch handles this issue.

4. page thrash handled ineffectively.

The fix is: on page thrash detection shutdown readahead.

attached patch 4_pagethrash.patch handles this issue.

5. slow read path is too slow.

I could not figure out a way to atleast-read-the-requested-
number-of-pages if readahead is shutdown, without incorporating
the readsize parameter to page_cache_readahead(). So had
to pick some of your code in filemap.c to do that. Thanks!

attached patch 5_fixedslowread.patch handles this issue.


Apart from this you have noticed other issues

6. cache lookup done unneccessrily twice for pagecache_hits.

I have not handled this issue currently. But should be doable
if I introducing a flag, which notes when readahead is
shutdown by pagecahche hits. And hence attempts to lookup
the page only once.


And you have other features in your patch which will be the real
differentiating factors.

7. exponential expand and shrink of window sizes.

8. overlapped read of current window and ahead window.

( I think both are desirable feature )

I did run some premilinary tests using your patch and the above patches
and found

your patch was doing slightly better on iozone and sysbench.
however the above patch were doing slightly better with DSS workload.

But my setup is rather tiny compared to your setup, so my comparison
is rather incomplete.

RP
--- tmp/mm/readahead.c 2004-09-29 17:45:02.120703232 -0700
+++ linux-2.6.8/mm/readahead.c 2004-09-29 17:45:12.043194784 -0700
@@ -139,8 +139,7 @@ out:
* start: Page index at which we started the readahead
* size: Number of pages in that read
* Together, these form the "current window".
- * Together, start and size represent the `readahead window'.
- * next_size: The number of pages to read on the next readahead miss.
+ * next_size: The number of pages to read in the next "ahead window".
* Has the magical value -1UL if readahead has been disabled.
* prev_page: The page which the readahead algorithm most-recently inspected.
* prev_page is mainly an optimisation: if page_cache_readahead
@@ -149,6 +148,12 @@ out:
* changes.
* ahead_start,
* ahead_size: Together, these form the "ahead window".
+ * currnt_wnd_hit: The average number of pages accessed in current
+ * 'current window'
+ * average: Tracks the average linearity of read requests. This value is
+ * used to (a) determine if "ahead window" has to be read in or not
+ * (b) determine the number of pages to read into the new
+ * "current window"
* ra_pages: The externally controlled max readahead for this fd.
*
* When readahead is in the "maximally shrunk" state (next_size == -1UL),
@@ -159,9 +164,9 @@ out:
*
* The readahead code manages two windows - the "current" and the "ahead"
* windows. The intent is that while the application is walking the pages
- * in the current window, I/O is underway on the ahead window. When the
- * current window is fully traversed, it is replaced by the ahead window
- * and the ahead window is invalidated. When this copying happens, the
+ * in the "current window", I/O is underway on the "ahead window". When the
+ * "current window" is fully traversed, it is replaced by the "ahead window"
+ * and the "ahead window" is invalidated. When this copying happens, the
* new current window's pages are probably still locked. When I/O has
* completed, we submit a new batch of I/O, creating a new ahead window.
*
@@ -171,11 +176,12 @@ out:
* ^start ^start+size
* ^ahead_start ^ahead_start+ahead_size
*
- * ^ When this page is read, we submit I/O for the
- * ahead window.
+ * ^ we submit I/O for the ahead window only if the reads
+ * in the past are sufficiently linear with an average
+ * linearity that is greater than the max readahead size.
*
* A `readahead hit' occurs when a read request is made against a page which is
- * inside the current window. Hits are good, and the window size (next_size)
+ * inside the "current window". Hits are good, and the window size (next_size)
* is grown aggressively when hits occur. Two pages are added to the next
* window size on each hit, which will end up doubling the next window size by
* the time I/O is submitted for it.
@@ -183,18 +189,17 @@ out:
* If readahead hits are more sparse (say, the application is only reading
* every second page) then the window will build more slowly.
*
- * On a readahead miss (the application seeked away) the readahead window is
- * shrunk by 25%. We don't want to drop it too aggressively, because it is a
- * good assumption that an application which has built a good readahead window
- * will continue to perform linear reads. Either at the new file position, or
- * at the old one after another seek.
+ * On a "current window" miss (the application seeked away) the readahead
+ * window is shrunk by 2 pages. We don't want to drop it too aggressively,
+ * because it is a good assumption that an application which has built a
+ * good readahead window will continue to perform linear reads. Either
+ * at the new file position, or at the old one after another seek.
*
* After enough misses, readahead is fully disabled. (next_size = -1UL).
*
- * There is a special-case: if the first page which the application tries to
- * read happens to be the first page of the file, it is assumed that a linear
- * read is about to happen and the window is immediately set to half of the
- * device maximum.
+ * There is a special-case: Irrespective of which page in the file is accessed
+ * the first time, it is assumed that a linear read is about to happen and the
+ * window is immediately set to half of the max readahead size.
*
* A page request at (start + size) is not a miss at all - it's just a part of
* sequential file reading.
@@ -443,9 +448,12 @@ page_cache_readahead(struct address_spac
/*
* Control now returns, probably to sleep until I/O
* completes against the first ahead page.
- * When the second page in the old ahead window is
- * requested, control will return here and more I/O
- * will be submitted to build the new ahead window.
+ * When some page in the old "ahead window" (which is
+ * now the "current window") is requested, more I/O
+ * will be submitted to build the new ahead window,
+ * provided the past read patterns have demonstrated
+ * that the pages in the ahead window are not wasted
+ * exessively.
*/
goto out;
}
@@ -460,7 +468,8 @@ do_io:
/* Heuristic: there is a high probability
* that around ra->average number of
* pages shall be accessed in the next
- * current window.
+ * current window. So we just bring in
+ * that many pages.
*/
average = ra->average;
if (ra->currnt_wnd_hit > average)
@@ -492,7 +501,7 @@ do_io:
/*
* If the average io-size is more than maximum
* readahead size of the file the io pattern is
- * sequential. Hence bring in the readahead window
+ * linear. Hence bring in the readahead window
* immediately.
* If the average io-size is less than maximum
* readahead size of the file the io pattern is--- linux-2.6.8/mm/readahead.c 2004-09-29 17:47:02.677375832 -0700
+++ linux-2.6.8.comment/mm/readahead.c 2004-09-29 17:39:00.880620072 -0700
@@ -329,18 +329,16 @@ int do_page_cache_readahead(struct addre
* occasionally.
*/
static inline void
-check_ra_success(struct file_ra_state *ra, pgoff_t attempt,
- pgoff_t actual, pgoff_t orig_next_size)
+check_ra_success(struct file_ra_state *ra, pgoff_t attempt, pgoff_t actual)
{
- if (actual == 0) {
- if (orig_next_size > 1) {
- ra->next_size = orig_next_size - 1;
- if (ra->ahead_size)
- ra->ahead_size = ra->next_size;
- } else {
- ra->next_size = -1UL;
- ra->size = 0;
- }
+ /*
+ * decrement window size by the number of pages hitting
+ * the page cache
+ */
+ ra->next_size -= attempt - actual;
+ if (ra->next_size <= 0) {
+ ra->next_size = -1UL;
+ ra->size = 0;
}
}

@@ -353,7 +351,6 @@ page_cache_readahead(struct address_spac
struct file *filp, unsigned long offset)
{
unsigned max;
- unsigned orig_next_size;
unsigned actual;
int first_access=0;
unsigned long average;
@@ -377,7 +374,6 @@ page_cache_readahead(struct address_spac
if (max == 0)
goto out; /* No readahead */

- orig_next_size = ra->next_size;

if (ra->next_size == 0) {
/*
@@ -489,7 +485,7 @@ do_io:
* time, the ahead window might get closed if all
* the pages are already in the cache.
*/
- check_ra_success(ra, ra->size, actual, orig_next_size);
+ check_ra_success(ra, ra->size, actual);
}
} else {
/*
@@ -516,8 +512,7 @@ do_io:
ra->ahead_size = ra->next_size;
actual = do_page_cache_readahead(mapping, filp,
ra->ahead_start, ra->ahead_size);
- check_ra_success(ra, ra->ahead_size,
- actual, orig_next_size);
+ check_ra_success(ra, ra->ahead_size, actual);
}
}
}--- linux-2.6.8/mm/readahead.c 2004-09-29 17:48:17.780958352 -0700
+++ linux-2.6.8.comment.pagecachehit/mm/readahead.c 2004-09-29 17:52:49.299681232 -0700
@@ -314,10 +314,11 @@ int force_page_cache_readahead(struct ad
int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
unsigned long offset, unsigned long nr_to_read)
{
- if (!bdi_read_congested(mapping->backing_dev_info))
- return __do_page_cache_readahead(mapping, filp,
- offset, nr_to_read);
- return 0;
+ if (bdi_read_congested(mapping->backing_dev_info))
+ return -BDI_read_congested;
+
+ return __do_page_cache_readahead(mapping, filp,
+ offset, nr_to_read);
}

/*
@@ -477,16 +478,9 @@ do_io:
ra->size = ra->next_size;
ra->ahead_start = 0; /* Invalidate these */
ra->ahead_size = 0;
- actual = do_page_cache_readahead(mapping, filp, offset,
+ actual = force_page_cache_readahead(mapping, filp, offset,
ra->size);
- if(!first_access) {
- /*
- * do not adjust the readahead window size the first
- * time, the ahead window might get closed if all
- * the pages are already in the cache.
- */
- check_ra_success(ra, ra->size, actual);
- }
+ check_ra_success(ra, ra->size, actual);
} else {
/*
* This read request is within the current window. It may be
@@ -512,7 +506,17 @@ do_io:
ra->ahead_size = ra->next_size;
actual = do_page_cache_readahead(mapping, filp,
ra->ahead_start, ra->ahead_size);
- check_ra_success(ra, ra->ahead_size, actual);
+ if (actual == -BDI_read_congested) {
+ /* queue is congested and no pages
+ * are read. So just behave as if
+ * readahead was not attempted at all.
+ */
+ ra->ahead_start = 0;
+ ra->ahead_size = 0;
+ } else {
+ check_ra_success(ra, ra->ahead_size,
+ actual);
+ }
}
}
}--- linux-2.6.8/mm/readahead.c 2004-09-29 17:53:37.120411376 -0700
+++ linux-2.6.8.comment.pagecachehit.queuecongestion/mm/readahead.c 2004-09-29 17:55:47.662565952 -0700
@@ -535,10 +535,8 @@ out:
* if we are getting misses against sequential file offsets. If so, and this
* persists then resume readahead.
*
- * Otherwise we're thrashing, so shrink the readahead window by three pages.
- * This is because it is grown by two pages on a readahead hit. Theory being
- * that the readahead window size will stabilise around the maximum level at
- * which there is no thrashing.
+ * Otherwise we're thrashing, close the readahead window. There is no point
+ * reading ahead if page-thrashing is noticed.
*/
void handle_ra_miss(struct address_space *mapping,
struct file_ra_state *ra, pgoff_t offset)
@@ -561,11 +559,8 @@ void handle_ra_miss(struct address_space
}
ra->prev_page = offset;
} else {
- const unsigned long min = get_min_readahead(ra);
-
- ra->next_size -= 3;
- if (ra->next_size < min)
- ra->next_size = min;
+ ra->next_size = -1UL;
+ ra->size = 0;
}
}
--- linux-2.6.8/mm/readahead.c 2004-09-29 17:57:00.133548688 -0700
+++ linux-2.6.8.comment.pagecachehit.queuecongestion.pagethrash/mm/readahead.c 2004-09-29 18:12:58.967783728 -0700
@@ -346,13 +346,19 @@ check_ra_success(struct file_ra_state *r
/*
* page_cache_readahead is the main function. If performs the adaptive
* readahead window size management and submits the readahead I/O.
+ *
+ * TOBEDONE: the size parameter is currently used to read the requested
+ * number of pages, only when the readahead is disabled. This parameter
+ * can be used to do more usefully in ramping up quickly readahead size
+ * as well as overlaping read of current window and readahead window.
*/
-void
+unsigned long
page_cache_readahead(struct address_space *mapping, struct file_ra_state *ra,
- struct file *filp, unsigned long offset)
+ struct file *filp, unsigned long offset,
+ unsigned long size)
{
- unsigned max;
- unsigned actual;
+ unsigned long max, min;
+ unsigned long actual;
int first_access=0;
unsigned long average;

@@ -375,6 +381,20 @@ page_cache_readahead(struct address_spac
if (max == 0)
goto out; /* No readahead */

+ if (ra->next_size == -1UL) {
+ unsigned long s_size;
+ /*
+ * Well readahead is disabled! So just read as many
+ * pages requested for now.
+ */
+ min = get_min_readahead(ra);
+ s_size = min((unsigned long)(max), (unsigned long)size);
+ handle_ra_miss(mapping, ra, offset, s_size);
+ (void )force_page_cache_readahead(mapping, filp,
+ offset, s_size);
+ return s_size;
+ }
+

if (ra->next_size == 0) {
/*
@@ -521,7 +541,10 @@ do_io:
}
}
out:
- return;
+ /*
+ * handled the request for page 'offset'. Hence return 1.
+ */
+ return 1;
}


@@ -538,26 +561,27 @@ out:
* Otherwise we're thrashing, close the readahead window. There is no point
* reading ahead if page-thrashing is noticed.
*/
-void handle_ra_miss(struct address_space *mapping,
- struct file_ra_state *ra, pgoff_t offset)
+void
+handle_ra_miss(struct address_space *mapping, struct file_ra_state *ra,
+ pgoff_t offset, unsigned long size)
{
if (ra->next_size == -1UL) {
- const unsigned long max = get_max_readahead(ra);
-
if (offset != ra->prev_page + 1) {
- ra->size = ra->size?ra->size-1:0; /* Not sequential */
+ ra->size += size - 2;
} else {
- ra->size++; /* A sequential read */
+ const unsigned long max = get_max_readahead(ra);
+ ra->size += size;
if (ra->size >= max) { /* Resume readahead */
- ra->start = offset - max;
+ ra->start = offset + size - max;
ra->next_size = max;
ra->size = max;
ra->ahead_start = 0;
ra->ahead_size = 0;
ra->average = max / 2;
+ ra->currnt_wnd_hit = 0;
}
}
- ra->prev_page = offset;
+ ra->prev_page = offset + size - 1;
} else {
ra->next_size = -1UL;
ra->size = 0;
\
 
 \ /
  Last update: 2005-03-22 14:06    [W:0.972 / U:0.256 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site