Hi Jens,
Patch 1 changes deadline to dispatch in a CSCAN-like manner for improved
throughput. This fixes two performance issues, namely deadline avalanche
that occurs under high load, and FCFS-like behaviour for mixed read/write
workloads. Benchmark results are below.
Patches 2 and 3 introduce some non-functional changes, mostly for style
and documentation.
Thanks,
-- Aaron
-----------------------------------x8-----------------------------------
Figures quoted in KiB/s unless otherwise noted. Disk under test is a 73GiB
15krpm Cheetah. Host machine is a 2-way IA64 w/ 2GiB RAM.
fio 16KiB reads
-------------------------------
Load Bandwidth %
New Old
32 3585.789 3580.412 0.1
64 3657.062 3361.460 8.8
128 3844.572 2459.811 56
256 3860.864 2455.446 57
512 3896.296 2472.468 58
fio 16KiB read/write (50/50)
-------------------------------
Load Bandwidth %
New Old
32 3461.336 2630.765 32
64 3715.625 2369.822 57
128 3931.525 2375.380 66
256 3953.333 2377.132 66
fio 1MiB reads
-------------------------------
Load Bandwidth %
New Old
32 49502.789 45744.023 8.2
64 50109.827 45776.416 9.5
128 50232.672 45884.900 9.5
256 50641.255 46266.082 9.5
fio 1MiB read/write (50/50)
-------------------------------
Load Bandwidth %
New Old
32 48543.272 44557.494 8.9
64 49388.251 44597.785 11
128 49794.080 45131.637 10
256 50632.356 45777.333 11
Postmark
-------------------------------
5 instances, 10k files, 100 dirs, 15k xacts, XFS
Old: 10.79 xact/s
New: 11.26 xact/s (+4.2%)
* convert goto to simpler while loop;
* use rq_end_sector() instead of computing manually;
* fix false comments;
* remove spurious whitespace;
* convert rq_rb_root macro to an inline function.
Signed-off-by: Aaron Carroll <[email protected]>
---
block/deadline-iosched.c | 26 +++++++++++++-------------
1 files changed, 13 insertions(+), 13 deletions(-)
diff --git a/block/deadline-iosched.c b/block/deadline-iosched.c
index 07b80e4..fd31117 100644
--- a/block/deadline-iosched.c
+++ b/block/deadline-iosched.c
@@ -33,7 +33,7 @@ struct deadline_data {
*/
struct rb_root sort_list[2];
struct list_head fifo_list[2];
-
+
/*
* next in sort order. read, write or both are NULL
*/
@@ -53,7 +53,11 @@ struct deadline_data {
static void deadline_move_request(struct deadline_data *, struct request *);
-#define RQ_RB_ROOT(dd, rq) (&(dd)->sort_list[rq_data_dir((rq))])
+static inline struct rb_root *
+deadline_rb_root(struct deadline_data *dd, struct request *rq)
+{
+ return &dd->sort_list[rq_data_dir(rq)];
+}
/*
* get the request after `rq' in sector-sorted order
@@ -72,15 +76,11 @@ deadline_latter_request(struct request *rq)
static void
deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
{
- struct rb_root *root = RQ_RB_ROOT(dd, rq);
+ struct rb_root *root = deadline_rb_root(dd, rq);
struct request *__alias;
-retry:
- __alias = elv_rb_add(root, rq);
- if (unlikely(__alias)) {
+ while (unlikely(__alias = elv_rb_add(root, rq)))
deadline_move_request(dd, __alias);
- goto retry;
- }
}
static inline void
@@ -91,7 +91,7 @@ deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
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);
+ elv_rb_del(deadline_rb_root(dd, rq), rq);
}
/*
@@ -106,7 +106,7 @@ deadline_add_request(struct request_queue *q, struct request *rq)
deadline_add_rq_rb(dd, rq);
/*
- * set expire time (only used for reads) and add to fifo list
+ * set expire time and add to fifo list
*/
rq_set_fifo_time(rq, jiffies + dd->fifo_expire[data_dir]);
list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
@@ -162,7 +162,7 @@ static void deadline_merged_request(struct request_queue *q,
* if the merge was a front merge, we need to reposition request
*/
if (type == ELEVATOR_FRONT_MERGE) {
- elv_rb_del(RQ_RB_ROOT(dd, req), req);
+ elv_rb_del(deadline_rb_root(dd, req), req);
deadline_add_rq_rb(dd, req);
}
}
@@ -212,7 +212,7 @@ deadline_move_request(struct deadline_data *dd, struct request *rq)
dd->next_rq[WRITE] = NULL;
dd->next_rq[data_dir] = deadline_latter_request(rq);
- dd->last_sector = rq->sector + rq->nr_sectors;
+ dd->last_sector = rq_end_sector(rq);
/*
* take it off the sort and fifo list, move
@@ -222,7 +222,7 @@ deadline_move_request(struct deadline_data *dd, struct request *rq)
}
/*
- * deadline_check_fifo returns 0 if there are no expired reads on the fifo,
+ * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
* 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
*/
static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
--
1.5.4.5
Deadline currently only batches sector-contiguous requests, so except
for a few circumstances (e.g. requests in a single direction), it is
essentially first come first served. This is bad for throughput, so
change it to CSCAN, which means requests in a batch do not need to be
sequential and are issued in increasing sector order.
Signed-off-by: Aaron Carroll <[email protected]>
---
block/deadline-iosched.c | 14 +++-----------
1 files changed, 3 insertions(+), 11 deletions(-)
diff --git a/block/deadline-iosched.c b/block/deadline-iosched.c
index 342448c..07b80e4 100644
--- a/block/deadline-iosched.c
+++ b/block/deadline-iosched.c
@@ -258,17 +258,9 @@ static int deadline_dispatch_requests(struct request_queue *q, int force)
else
rq = dd->next_rq[READ];
- if (rq) {
- /* we have a "next request" */
-
- if (dd->last_sector != rq->sector)
- /* end the batch on a non sequential request */
- dd->batching += dd->fifo_batch;
-
- if (dd->batching < dd->fifo_batch)
- /* we are still entitled to batch */
- goto dispatch_request;
- }
+ if (rq && dd->batching < dd->fifo_batch)
+ /* we have a next request are still entitled to batch */
+ goto dispatch_request;
/*
* at this point we are not running a batch. select the appropriate
--
1.5.4.5
Update the description of fifo_batch to match the current implementation,
and include a description of how to tune it.
Signed-off-by: Aaron Carroll <[email protected]>
---
Documentation/block/deadline-iosched.txt | 14 ++++++++++----
1 files changed, 10 insertions(+), 4 deletions(-)
diff --git a/Documentation/block/deadline-iosched.txt b/Documentation/block/deadline-iosched.txt
index c23cab1..7257676 100644
--- a/Documentation/block/deadline-iosched.txt
+++ b/Documentation/block/deadline-iosched.txt
@@ -30,12 +30,18 @@ write_expire (in ms)
Similar to read_expire mentioned above, but for writes.
-fifo_batch
+fifo_batch (number of requests)
----------
-When a read request expires its deadline, we must move some requests from
-the sorted io scheduler list to the block device dispatch queue. fifo_batch
-controls how many requests we move.
+Requests are grouped into ``batches'' of a particular data direction (read or
+write) which are serviced in increasing sector order. To limit extra seeking,
+deadline expiries are only checked between batches. fifo_batch controls the
+maximum number of requests per batch.
+
+This parameter tunes the balance between per-request latency and aggregate
+throughput. When low latency is the primary concern, smaller is better (where
+a value of 1 yields first-come first-served behaviour). Increasing fifo_batch
+generally improves throughput, at the cost of latency variation.
writes_starved (number of dispatches)
--
1.5.4.5
On Thu, Aug 14 2008, Aaron Carroll wrote:
>
> Hi Jens,
>
> Patch 1 changes deadline to dispatch in a CSCAN-like manner for improved
> throughput. This fixes two performance issues, namely deadline avalanche
> that occurs under high load, and FCFS-like behaviour for mixed read/write
> workloads. Benchmark results are below.
>
> Patches 2 and 3 introduce some non-functional changes, mostly for style
> and documentation.
Looks good, thanks Aaron!
--
Jens Axboe