Skip to contentSkip to conversation

dm: dm-mpath: Provide high-resolution timer to HST with bio-mpath

PATCH In review Gabriel Krisman Bertazi 2 participants 4 messages Apr 27, 2022 – May 9, 2022 Device mapper

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dm: dm-mpath: Provide high-resolution timer to HST with bio-mpath

The precision loss of reading IO start_time with jiffies_to_nsecs instead of using a high resolution timer degrades HST path prediction for BIO-based mpath on high load workloads.

Below, I show the utilization percentage of a 10 disk multipath with asymmetrical disk access cost, while being exercised by a randwrite FIO benchmark with high submission queue depth (depth=64). It is possible to see that the HST path selection degrades heavily for high-iops in BIO-mpath, underutilizing the slower paths way beyond expected. This seems to be caused by the start_time truncation, which makes some IO to seem much slower than they actually is. In this scenario ST outperforms HST for bio-mpath, but not for mq-mpath, which already uses ktime_get_ns().

The third column shows utilization with this patch applied. It is easy to see that now HST prediction is much closer to the ideal distribution (calculated considering the real cost of each path).

|     |   ST | HST (orig) | HST(ktime) | Best |
| sdd | 0.17 |       0.20 |       0.17 | 0.18 |
| sde | 0.17 |       0.20 |       0.17 | 0.18 |
| sdf | 0.17 |       0.20 |       0.17 | 0.18 |
| sdg | 0.06 |       0.00 |       0.06 | 0.04 |
| sdh | 0.03 |       0.00 |       0.03 | 0.02 |
| sdi | 0.03 |       0.00 |       0.03 | 0.02 |
| sdj | 0.02 |       0.00 |       0.01 | 0.01 |
| sdk | 0.02 |       0.00 |       0.01 | 0.01 |
| sdl | 0.17 |       0.20 |       0.17 | 0.18 |
| sdm | 0.17 |       0.20 |       0.17 | 0.18 |

This issue was originally discussed [1] when we first merged HST, and this patch was left as a low hanging fruit to be solved later. I don't think anyone is using HST with BIO mpath, but it'd be neat to get it sorted out.

Regarding the implementation, as suggested by Mike in that mail thread, in order to avoid the overhead of ktime_get_ns for other selectors, this patch adds a flag for the selector code to request the high-resolution timer.

I tested this using the same benchmark used in the original HST submission.

Full test and benchmark scripts are available here:

https://people.collabora.com/~krisman/HST-BIO-MPATH/

[1] https://lore.kernel.org/lkml/[email protected]/T/

---
 drivers/md/dm-mpath.c                      | 16 ++++++++++++++--
 drivers/md/dm-path-selector.h              | 13 +++++++++++++
 drivers/md/dm-ps-historical-service-time.c |  1 +
 3 files changed, 28 insertions(+), 2 deletions(-)
drivers/md/dm-mpath.c +14 -2
index f4719b65e5e3..c58cfcd87d04 100644
@@ -105,6 +105,7 @@ struct multipath {
 struct dm_mpath_io {
 	struct pgpath *pgpath;
 	size_t nr_bytes;
+	u64 start_time_ns;
 };
 
 typedef int (*action_fn) (struct pgpath *pgpath);
@@ -647,6 +648,11 @@ static int __multipath_map_bio(struct multipath *m, struct bio *bio,
 
 	mpio->pgpath = pgpath;
 
+	if (pgpath->pg->ps.type->flags & PS_NEED_HR_TIMER_FL)
+		mpio->start_time_ns = ktime_get_ns();
+	else
+		mpio->start_time_ns = 0;
+
 	bio->bi_status = 0;
 	bio_set_dev(bio, pgpath->path.dev->bdev);
 	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
@@ -1715,9 +1721,15 @@ static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
 	if (pgpath) {
 		struct path_selector *ps = &pgpath->pg->ps;
 
-		if (ps->type->end_io)
+		if (ps->type->end_io) {
+			u64 st_time = mpio->start_time_ns;
+
+			if (!st_time)
+				st_time = dm_start_time_ns_from_clone(clone);
+
 			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
-					 dm_start_time_ns_from_clone(clone));
+					 st_time);
+		}
 	}
 
 	return r;
drivers/md/dm-path-selector.h +13 -0
index c47bc0e20275..46710f364286 100644
@@ -26,6 +26,18 @@ struct path_selector {
 	void *context;
 };
 
+/* If a path selector uses this flag, a more precise timer
+ * (ktime_get_ns) is used to account for IO start time in BIO-based
+ * mpath.  This improves performance of some path selectors (i.e. HST),
+ * in exchange of a slightly higher overhead when submiting the
+ * BIO. The extra cost is usually offset by improved path selection for
+ * some benchmarks.
+ *
+ * This has no effect for request-based mpath, since it already uses a
+ * higher precision timer by default.
+ */
+#define PS_NEED_HR_TIMER_FL 0x1
+
 /* Information about a path selector type */
 struct path_selector_type {
 	char *name;
@@ -33,6 +45,7 @@ struct path_selector_type {
 
 	unsigned int table_args;
 	unsigned int info_args;
+	unsigned int flags;
 
 	/*
 	 * Constructs a path selector object, takes custom arguments
drivers/md/dm-ps-historical-service-time.c +1 -0
index 82f2a06153dc..d2233cc4ea97 100644
@@ -523,6 +523,7 @@ static int hst_end_io(struct path_selector *ps, struct dm_path *path,
 static struct path_selector_type hst_ps = {
 	.name		= "historical-service-time",
 	.module		= THIS_MODULE,
+	.flags		= PS_NEED_HR_TIMER_FL,
 	.table_args	= 1,
 	.info_args	= 3,
 	.create		= hst_create,
Signature · 6 lines
2.35.1

--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel
#1 lore raw
Quoted 41 lines · Gabriel Krisman Bertazi

The precision loss of reading IO start_time with jiffies_to_nsecs instead of using a high resolution timer degrades HST path prediction for BIO-based mpath on high load workloads.

Below, I show the utilization percentage of a 10 disk multipath with asymmetrical disk access cost, while being exercised by a randwrite FIO benchmark with high submission queue depth (depth=64). It is possible to see that the HST path selection degrades heavily for high-iops in BIO-mpath, underutilizing the slower paths way beyond expected. This seems to be caused by the start_time truncation, which makes some IO to seem much slower than they actually is. In this scenario ST outperforms HST for bio-mpath, but not for mq-mpath, which already uses ktime_get_ns().

The third column shows utilization with this patch applied. It is easy to see that now HST prediction is much closer to the ideal distribution (calculated considering the real cost of each path).

|     |   ST | HST (orig) | HST(ktime) | Best |
| sdd | 0.17 |       0.20 |       0.17 | 0.18 |
| sde | 0.17 |       0.20 |       0.17 | 0.18 |
| sdf | 0.17 |       0.20 |       0.17 | 0.18 |
| sdg | 0.06 |       0.00 |       0.06 | 0.04 |
| sdh | 0.03 |       0.00 |       0.03 | 0.02 |
| sdi | 0.03 |       0.00 |       0.03 | 0.02 |
| sdj | 0.02 |       0.00 |       0.01 | 0.01 |
| sdk | 0.02 |       0.00 |       0.01 | 0.01 |
| sdl | 0.17 |       0.20 |       0.17 | 0.18 |
| sdm | 0.17 |       0.20 |       0.17 | 0.18 |

This issue was originally discussed [1] when we first merged HST, and this patch was left as a low hanging fruit to be solved later. I don't think anyone is using HST with BIO mpath, but it'd be neat to get it sorted out.

Regarding the implementation, as suggested by Mike in that mail thread, in order to avoid the overhead of ktime_get_ns for other selectors, this patch adds a flag for the selector code to request the high-resolution timer.

I tested this using the same benchmark used in the original HST submission.

Quoted 8 lines

Full test and benchmark scripts are available here:

https://people.collabora.com/~krisman/HST-BIO-MPATH/

[1] https://lore.kernel.org/lkml/[email protected]/T/

Overall your code was OK, but I nudged it a bit further to be inkeeping with how 'features' flags have been implemented elsewhere (e.g. dm_target_type's features) -- by using a healer to test the flag, etc.

I also tweaked some other small implementation details. Please see:
https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-5.19&id=c06dfd124d46df9c482fbd1319b5fe19bcb1a110

--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel

#2 lore raw
Quoted 50 lines · Mike Snitzer
Quoted 47 lines · Gabriel Krisman Bertazi

The precision loss of reading IO start_time with jiffies_to_nsecs instead of using a high resolution timer degrades HST path prediction for BIO-based mpath on high load workloads.

Below, I show the utilization percentage of a 10 disk multipath with asymmetrical disk access cost, while being exercised by a randwrite FIO benchmark with high submission queue depth (depth=64). It is possible to see that the HST path selection degrades heavily for high-iops in BIO-mpath, underutilizing the slower paths way beyond expected. This seems to be caused by the start_time truncation, which makes some IO to seem much slower than they actually is. In this scenario ST outperforms HST for bio-mpath, but not for mq-mpath, which already uses ktime_get_ns().

The third column shows utilization with this patch applied. It is easy to see that now HST prediction is much closer to the ideal distribution (calculated considering the real cost of each path).

|     |   ST | HST (orig) | HST(ktime) | Best |
| sdd | 0.17 |       0.20 |       0.17 | 0.18 |
| sde | 0.17 |       0.20 |       0.17 | 0.18 |
| sdf | 0.17 |       0.20 |       0.17 | 0.18 |
| sdg | 0.06 |       0.00 |       0.06 | 0.04 |
| sdh | 0.03 |       0.00 |       0.03 | 0.02 |
| sdi | 0.03 |       0.00 |       0.03 | 0.02 |
| sdj | 0.02 |       0.00 |       0.01 | 0.01 |
| sdk | 0.02 |       0.00 |       0.01 | 0.01 |
| sdl | 0.17 |       0.20 |       0.17 | 0.18 |
| sdm | 0.17 |       0.20 |       0.17 | 0.18 |

This issue was originally discussed [1] when we first merged HST, and this patch was left as a low hanging fruit to be solved later. I don't think anyone is using HST with BIO mpath, but it'd be neat to get it sorted out.

Regarding the implementation, as suggested by Mike in that mail thread, in order to avoid the overhead of ktime_get_ns for other selectors, this patch adds a flag for the selector code to request the high-resolution timer.

I tested this using the same benchmark used in the original HST submission.

Full test and benchmark scripts are available here:

https://people.collabora.com/~krisman/HST-BIO-MPATH/

[1] https://lore.kernel.org/lkml/[email protected]/T/

Quoted 7 lines
Quoted 2 lines

Overall your code was OK, but I nudged it a bit further to be inkeeping with how 'features' flags have been implemented elsewhere (e.g. dm_target_type's features) -- by using a healer to test the flag, etc.

s/healer/helper/ ;)

--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel

#3 lore raw
In reply to Mike Snitzer
Quoted 7 lines · Mike Snitzer

Overall your code was OK, but I nudged it a bit further to be inkeeping with how 'features' flags have been implemented elsewhere (e.g. dm_target_type's features) -- by using a healer to test the flag, etc.

I also tweaked some other small implementation details. Please see:
https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-5.19&id=c06dfd124d46df9c482fbd1319b5fe19bcb1a110

looks good to me. thanks for the fixups.

Signature · 6 lines
Gabriel Krisman Bertazi

--
dm-devel mailing list
[email protected]
https://listman.redhat.com/mailman/listinfo/dm-devel