2007-10-30 04:03:36

by Aaron Carroll

[permalink] [raw]
Subject: [PATCH 0/3] Deadline iosched: Fix batching algorithm

Hi Jens,

The following patches correct some issues with the deadline I/O
scheduler and its batching algorithm.

Patch 1 is a simple function factorisation.
Patch 2 fixes a missing batch count reset, making the behaviour
closer to that implied by the documentation.
Patch 3 changes batch start points to resolve a disparity in
latency and bandwidth between high- and low-sector requests.


Thanks,
-- Aaron



2007-10-30 04:05:26

by Aaron Carroll

[permalink] [raw]
Subject: [PATCH 1/3] Deadline iosched: Factor out finding latter request

Factor finding the next request in sector-sorted order into
a function deadline_latter_request.

Signed-off-by: Aaron Carroll <[email protected]>
---
block/deadline-iosched.c | 28 +++++++++++++++++-----------
1 files changed, 17 insertions(+), 11 deletions(-)

diff --git a/block/deadline-iosched.c b/block/deadline-iosched.c
index 1a511ff..a44437e 100644
--- a/block/deadline-iosched.c
+++ b/block/deadline-iosched.c
@@ -55,6 +55,20 @@ static void deadline_move_request(struct deadline_data *, struct request *);

#define RQ_RB_ROOT(dd, rq) (&(dd)->sort_list[rq_data_dir((rq))])

+/*
+ * get the request after `rq' in sector-sorted order
+ */
+static inline struct request *
+deadline_latter_request(struct request *rq)
+{
+ struct rb_node *node = rb_next(&rq->rb_node);
+
+ if (node)
+ return rb_entry_rq(node);
+
+ return NULL;
+}
+
static void
deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
{
@@ -74,13 +88,8 @@ deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
{
const int data_dir = rq_data_dir(rq);

- if (dd->next_rq[data_dir] == rq) {
- struct rb_node *rbnext = rb_next(&rq->rb_node);
-
- dd->next_rq[data_dir] = NULL;
- if (rbnext)
- dd->next_rq[data_dir] = rb_entry_rq(rbnext);
- }
+ if (dd->next_rq[data_dir] == rq)
+ dd->next_rq[data_dir] = deadline_latter_request(rq);

elv_rb_del(RQ_RB_ROOT(dd, rq), rq);
}
@@ -198,14 +207,11 @@ static void
deadline_move_request(struct deadline_data *dd, struct request *rq)
{
const int data_dir = rq_data_dir(rq);
- struct rb_node *rbnext = rb_next(&rq->rb_node);

dd->next_rq[READ] = NULL;
dd->next_rq[WRITE] = NULL;
+ dd->next_rq[data_dir] = deadline_latter_request(rq);

- if (rbnext)
- dd->next_rq[data_dir] = rb_entry_rq(rbnext);
-
dd->last_sector = rq->sector + rq->nr_sectors;

/*

2007-10-30 04:13:25

by Aaron Carroll

[permalink] [raw]
Subject: [PATCH 3/3] Deadline iosched: Fix batching fairness

After switching data directions, deadline always starts the next batch
from the lowest-sector request. This gives excessive deadline expiries
and large latency and throughput disparity between high- and low-sector
requests; an order of magnitude in some tests.

This patch changes the batching behaviour so new batches start from the
request whose expiry is earliest.

Signed-off-by: Aaron Carroll <[email protected]>
---
block/deadline-iosched.c | 21 +++++++--------------
1 files changed, 7 insertions(+), 14 deletions(-)

diff --git a/block/deadline-iosched.c b/block/deadline-iosched.c
index cb94c83..a054eef 100644
--- a/block/deadline-iosched.c
+++ b/block/deadline-iosched.c
@@ -306,27 +306,20 @@ dispatch_writes:
dispatch_find_request:
/*
* we are not running a batch, find best request for selected data_dir
- * and start a new batch
*/
- if (deadline_check_fifo(dd, data_dir)) {
- /* An expired request exists - satisfy it */
+ if (deadline_check_fifo(dd, data_dir) || !dd->next_rq[data_dir]) {
+ /*
+ * A deadline has expired, the last request was in the other
+ * direction, or we have run out of higher-sectored requests.
+ * Start again from the request with the earliest expiry time.
+ */
rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
- } else if (dd->next_rq[data_dir]) {
+ } else {
/*
* The last req was the same dir and we have a next request in
* sort order. No expired requests so continue on from here.
*/
rq = dd->next_rq[data_dir];
- } else {
- struct rb_node *node;
- /*
- * The last req was the other direction or we have run out of
- * higher-sectored requests. Go back to the lowest sectored
- * request (1 way elevator) and start a new batch.
- */
- node = rb_first(&dd->sort_list[data_dir]);
- if (node)
- rq = rb_entry_rq(node);
}

dd->batching = 0;

2007-10-30 04:13:38

by Aaron Carroll

[permalink] [raw]
Subject: [PATCH 2/3] Deadline iosched: Reset batch for ordered requests

The deadline I/O scheduler does not reset the batch count when starting
a new batch at a higher-sectored request. This means the second and
subsequent batch in the same data direction will never exceed a single
request in size whenever higher-sectored requests are pending.

This patch gives new batches in the same data direction as old ones
their full quota of requests by resetting the batch count.

Signed-off-by: Aaron Carroll <[email protected]>
---
block/deadline-iosched.c | 6 +++---
1 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/block/deadline-iosched.c b/block/deadline-iosched.c
index a44437e..cb94c83 100644
--- a/block/deadline-iosched.c
+++ b/block/deadline-iosched.c
@@ -306,12 +306,11 @@ dispatch_writes:
dispatch_find_request:
/*
* we are not running a batch, find best request for selected data_dir
+ * and start a new batch
*/
if (deadline_check_fifo(dd, data_dir)) {
/* An expired request exists - satisfy it */
- dd->batching = 0;
rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
-
} else if (dd->next_rq[data_dir]) {
/*
* The last req was the same dir and we have a next request in
@@ -325,12 +324,13 @@ dispatch_find_request:
* higher-sectored requests. Go back to the lowest sectored
* request (1 way elevator) and start a new batch.
*/
- dd->batching = 0;
node = rb_first(&dd->sort_list[data_dir]);
if (node)
rq = rb_entry_rq(node);
}

+ dd->batching = 0;
+
dispatch_request:
/*
* rq is the selected appropriate request.

2007-10-30 09:43:20

by Jens Axboe

[permalink] [raw]
Subject: Re: [PATCH 0/3] Deadline iosched: Fix batching algorithm

On Tue, Oct 30 2007, Aaron Carroll wrote:
> Hi Jens,
>
> The following patches correct some issues with the deadline I/O
> scheduler and its batching algorithm.
>
> Patch 1 is a simple function factorisation.
> Patch 2 fixes a missing batch count reset, making the behaviour
> closer to that implied by the documentation.

1+2 are no brainers, applied.

> Patch 3 changes batch start points to resolve a disparity in
> latency and bandwidth between high- and low-sector requests.

This one makes a lot of sense, applied as well.

Thanks!

--
Jens Axboe