dm: dm-mpath: Provide high-resolution timer to HST with bio-mpath
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
@@ -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
@@ -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
@@ -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
Quoted 41 lines · Gabriel Krisman Bertazi
The precision loss of reading IO
start_timewithjiffies_to_nsecsinstead 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_timetruncation, 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 usesktime_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_nsfor 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/
Signed-off-by Gabriel Krisman Bertazi <[email protected]>
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
Quoted 50 lines · Mike Snitzer
Quoted 47 lines · Gabriel Krisman Bertazi
The precision loss of reading IO
start_timewithjiffies_to_nsecsinstead 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_timetruncation, 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 usesktime_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_nsfor 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:
Quoted 7 lines
Quoted 2 lines
Signed-off-by Gabriel Krisman Bertazi <[email protected]>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
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