Skip to contentSkip to conversation

optimize raid1 read balance for SSD

PATCH Superseded Shaohua Li 4 participants 10 messages Jun 13, 2012 – Jun 28, 2012 MD RAID

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Shaohua Li

author
#0 lore raw

v2 cover optimize raid1 read balance for SSD

raid1 read balance is an important algorithm to make read performance optimal. It's a distance based algorithm, eg, for each request dispatch, choose disk whose last finished request is close the request. This is great for hard disk. But SSD has some special characteristics:

1. nonrotational. Distance means nothing for SSD, though merging small rquests to big request is still optimal for SSD. If no merge, distributing rquests to raid disks as more as possible is more optimal.

2. Getting too big request isn't always optimal. For hard disk, compared to spindle move, data transfer overhead is trival, so we always prefer bigger request. In SSD, request size exceeds specific value, performance isn't always increased with request size increased. An example is readahead. If readahead merges too big request and causes some disks idle, the performance is less optimal than that when all disks are busy and running small requests.

The patches try to address the issues. The first two patches are clean up. The third patch addresses the first item above. The forth addresses the second item above. The idea can be applied to raid10 too, which is in my todo list.

V1->V2:
rebase to latest kernel.

author
#1 lore raw

PATCH v2 1/4 raid1: move distance based read balance to a separate function

Move distance based read balance algorithm to a separate function. No functional change.

Signed-off-by Shaohua Li <[email protected]>
---
 drivers/md/raid1.c |   42 ++++++++++++++++++++++++++++++------------
 1 file changed, 30 insertions(+), 12 deletions(-)
drivers/md/raid1.c +30 -12
Index: linux/drivers/md/raid1.c
===================================================================
@@ -463,6 +463,31 @@ static void raid1_end_write_request(stru
 		bio_put(to_put);
 }
 
+static int read_balance_measure_distance(struct r1conf *conf,
+	struct r1bio *r1_bio, int disk, int *best_disk, sector_t *best_dist)
+{
+	const sector_t this_sector = r1_bio->sector;
+	struct md_rdev *rdev;
+	sector_t dist;
+
+	rdev = rcu_dereference(conf->mirrors[disk].rdev);
+
+	dist = abs(this_sector - conf->mirrors[disk].head_position);
+	/* Don't change to another disk for sequential reads */
+	if (conf->next_seq_sect == this_sector
+	    || dist == 0
+	    /* If device is idle, use it */
+	    || atomic_read(&rdev->nr_pending) == 0) {
+		*best_disk = disk;
+		return 0;
+	}
+
+	if (dist < *best_dist) {
+		*best_dist = dist;
+		*best_disk = disk;
+	}
+	return 1;
+}
 
 /*
  * This routine returns the disk from which the requested read should

@@ -512,7 +537,6 @@ static int read_balance(struct r1conf *c
 	}
 
 	for (i = 0 ; i < conf->raid_disks * 2 ; i++) {
-		sector_t dist;
 		sector_t first_bad;
 		int bad_sectors;
 

@@ -577,20 +601,14 @@ static int read_balance(struct r1conf *c
 		} else
 			best_good_sectors = sectors;
 
-		dist = abs(this_sector - conf->mirrors[disk].head_position);
-		if (choose_first
-		    /* Don't change to another disk for sequential reads */
-		    || conf->next_seq_sect == this_sector
-		    || dist == 0
-		    /* If device is idle, use it */
-		    || atomic_read(&rdev->nr_pending) == 0) {
+		if (choose_first) {
 			best_disk = disk;
 			break;
 		}
-		if (dist < best_dist) {
-			best_dist = dist;
-			best_disk = disk;
-		}
+
+		if (!read_balance_measure_distance(conf, r1_bio, disk,
+		    &best_disk, &best_dist))
+			break;
 	}
 
 	if (best_disk >= 0) {
author
#2 lore raw

PATCH v2 2/4 raid1: make sequential read detection per disk based

In reply to Shaohua Li

Currently the sequential read detection is global wide. It's natural to make it per disk based, which can improve the detection for concurrent multiple sequential reads. And next patch will make SSD read balance not use distance based algorithm, where this change help detect truly sequential read for SSD.

Signed-off-by Shaohua Li <[email protected]>
---
 drivers/md/raid1.c |   29 ++++++-----------------------
 drivers/md/raid1.h |   11 +++++------
 2 files changed, 11 insertions(+), 29 deletions(-)
drivers/md/raid1.c +6 -23
Index: linux/drivers/md/raid1.c
===================================================================
@@ -474,7 +474,7 @@ static int read_balance_measure_distance
 
 	dist = abs(this_sector - conf->mirrors[disk].head_position);
 	/* Don't change to another disk for sequential reads */
-	if (conf->next_seq_sect == this_sector
+	if (conf->mirrors[disk].next_seq_sect == this_sector
 	    || dist == 0
 	    /* If device is idle, use it */
 	    || atomic_read(&rdev->nr_pending) == 0) {

@@ -508,7 +508,6 @@ static int read_balance(struct r1conf *c
 	const sector_t this_sector = r1_bio->sector;
 	int sectors;
 	int best_good_sectors;
-	int start_disk;
 	int best_disk;
 	int i;
 	sector_t best_dist;

@@ -528,19 +527,16 @@ static int read_balance(struct r1conf *c
 	best_good_sectors = 0;
 
 	if (conf->mddev->recovery_cp < MaxSector &&
-	    (this_sector + sectors >= conf->next_resync)) {
+	    (this_sector + sectors >= conf->next_resync))
 		choose_first = 1;
-		start_disk = 0;
-	} else {
+	else
 		choose_first = 0;
-		start_disk = conf->last_used;
-	}
 
 	for (i = 0 ; i < conf->raid_disks * 2 ; i++) {
 		sector_t first_bad;
 		int bad_sectors;
 
-		int disk = start_disk + i;
+		int disk = i;
 		if (disk >= conf->raid_disks)
 			disk -= conf->raid_disks;
 

@@ -624,8 +620,7 @@ static int read_balance(struct r1conf *c
 			goto retry;
 		}
 		sectors = best_good_sectors;
-		conf->next_seq_sect = this_sector + sectors;
-		conf->last_used = best_disk;
+		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 	}
 	rcu_read_unlock();
 	*max_sectors = sectors;

@@ -2599,7 +2594,6 @@ static struct r1conf *setup_conf(struct
 	conf->recovery_disabled = mddev->recovery_disabled - 1;
 
 	err = -EIO;
-	conf->last_used = -1;
 	for (i = 0; i < conf->raid_disks * 2; i++) {
 
 		disk = conf->mirrors + i;

@@ -2625,19 +2619,9 @@ static struct r1conf *setup_conf(struct
 			if (disk->rdev &&
 			    (disk->rdev->saved_raid_disk < 0))
 				conf->fullsync = 1;
-		} else if (conf->last_used < 0)
-			/*
-			 * The first working device is used as a
-			 * starting point to read balancing.
-			 */
-			conf->last_used = i;
+		}
 	}
 
-	if (conf->last_used < 0) {
-		printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
-		       mdname(mddev));
-		goto abort;
-	}
 	err = -ENOMEM;
 	conf->thread = md_register_thread(raid1d, mddev, NULL);
 	if (!conf->thread) {

@@ -2894,7 +2878,6 @@ static int raid1_reshape(struct mddev *m
 	conf->raid_disks = mddev->raid_disks = raid_disks;
 	mddev->delta_disks = 0;
 
-	conf->last_used = 0; /* just make sure it is in-range */
 	lower_barrier(conf);
 
 	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
drivers/md/raid1.h +5 -6
Index: linux/drivers/md/raid1.h
===================================================================
@@ -4,6 +4,11 @@
 struct mirror_info {
 	struct md_rdev	*rdev;
 	sector_t	head_position;
+
+	/* When choose the best device for a read (read_balance())
+	 * we try to keep sequential reads one the same device
+	 */
+	sector_t	next_seq_sect;
 };
 
 /*

@@ -29,12 +34,6 @@ struct r1conf {
 						 */
 	int			raid_disks;
 
-	/* When choose the best device for a read (read_balance())
-	 * we try to keep sequential reads one the same device
-	 * using 'last_used' and 'next_seq_sect'
-	 */
-	int			last_used;
-	sector_t		next_seq_sect;
 	/* During resync, read_balancing is only allowed on the part
 	 * of the array that has been resynced.  'next_resync' tells us
 	 * where that is.
author
#3 lore raw

PATCH v2 3/4 raid1: read balance chooses idlest disk

In reply to Shaohua Li

SSD hasn't spindle, distance between requests means nothing. And the original distance based algorithm sometimes can cause severe performance issue for SSD raid.

Considering two thread groups, one accesses file A, the other access file B. The first group will access one disk and the second will access the other disk, because requests are near from one group and far between groups. In this case, read balance might keep one disk very busy but the other relative idle. For SSD, we should try best to distribute requests to as more disks as possible. There isn't spindle move penality anyway.

With below patch, I can see more than 50% throughput improvement sometimes depending on workloads.

The only exception is small requests can be merged to a big request which typically can drive higher throughput for SSD too. Such small requests are sequential reads. Unlike hard disk, sequential read which can't be merged (for example direct IO, or read without readahead) can be ignored for SSD. Again there is no spindle move penality. readahead dispatches small requests and such requests can be merged.

Last patch can help detect sequential read well, at least if concurrent read number isn't greater than raid disk number. In that case, distance based algorithm doesn't work well too.

Signed-off-by Shaohua Li <[email protected]>
---
 drivers/md/raid1.c |   55 ++++++++++++++++++++++++++++++++++++++++++++++++++---
 drivers/md/raid1.h |    2 +
 2 files changed, 54 insertions(+), 3 deletions(-)
drivers/md/raid1.c +52 -3
Index: linux/drivers/md/raid1.c
===================================================================
@@ -463,6 +463,43 @@ static void raid1_end_write_request(stru
 		bio_put(to_put);
 }
 
+static int read_balance_measure_ssd(struct r1conf *conf, struct r1bio *r1_bio,
+	int disk, int *best_disk, unsigned int *min_pending)
+{
+	const sector_t this_sector = r1_bio->sector;
+	struct md_rdev *rdev;
+	unsigned int pending;
+
+	rdev = rcu_dereference(conf->mirrors[disk].rdev);
+	pending = atomic_read(&rdev->nr_pending);
+
+	/* big request IO helps SSD too, allow sequential IO merge */
+	if (conf->mirrors[disk].next_seq_sect == this_sector) {
+		sector_t dist;
+		dist = abs(this_sector - conf->mirrors[disk].head_position);
+		/*
+		 * head_position is for finished request, such reqeust can't be
+		 * merged with current request, so it means nothing for SSD
+		 */
+		if (dist != 0)
+			goto done;
+	}
+
+	/* If device is idle, use it */
+	if (pending == 0)
+		goto done;
+
+	/* find device with less requests pending */
+	if (*min_pending > pending) {
+		*min_pending = pending;
+		*best_disk = disk;
+	}
+	return 1;
+done:
+	*best_disk = disk;
+	return 0;
+}
+
 static int read_balance_measure_distance(struct r1conf *conf,
 	struct r1bio *r1_bio, int disk, int *best_disk, sector_t *best_dist)
 {

@@ -511,6 +548,7 @@ static int read_balance(struct r1conf *c
 	int best_disk;
 	int i;
 	sector_t best_dist;
+	unsigned int min_pending;
 	struct md_rdev *rdev;
 	int choose_first;
 

@@ -524,6 +562,7 @@ static int read_balance(struct r1conf *c
 	sectors = r1_bio->sectors;
 	best_disk = -1;
 	best_dist = MaxSector;
+	min_pending = -1;
 	best_good_sectors = 0;
 
 	if (conf->mddev->recovery_cp < MaxSector &&

@@ -602,9 +641,15 @@ static int read_balance(struct r1conf *c
 			break;
 		}
 
-		if (!read_balance_measure_distance(conf, r1_bio, disk,
-		    &best_disk, &best_dist))
-			break;
+		if (!conf->nonrotational) {
+			if (!read_balance_measure_distance(conf, r1_bio, disk,
+			    &best_disk, &best_dist))
+				break;
+		} else {
+			if (!read_balance_measure_ssd(conf, r1_bio, disk,
+			    &best_disk, &min_pending))
+				break;
+		}
 	}
 
 	if (best_disk >= 0) {

@@ -2533,6 +2578,7 @@ static struct r1conf *setup_conf(struct
 	struct mirror_info *disk;
 	struct md_rdev *rdev;
 	int err = -ENOMEM;
+	bool nonrotational = true;
 
 	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
 	if (!conf)

@@ -2581,7 +2627,10 @@ static struct r1conf *setup_conf(struct
 			mddev->merge_check_needed = 1;
 
 		disk->head_position = 0;
+		if (!blk_queue_nonrot(bdev_get_queue(rdev->bdev)))
+			nonrotational = false;
 	}
+	conf->nonrotational = nonrotational;
 	conf->raid_disks = mddev->raid_disks;
 	conf->mddev = mddev;
 	INIT_LIST_HEAD(&conf->retry_list);
drivers/md/raid1.h +2 -0
Index: linux/drivers/md/raid1.h
===================================================================
@@ -65,6 +65,8 @@ struct r1conf {
 	int			nr_queued;
 	int			barrier;
 
+	int			nonrotational;
+
 	/* Set to 1 if a full sync is needed, (fresh device added).
 	 * Cleared when a sync completes.
 	 */
author
#4 lore raw

PATCH v2 4/4 raid1: split large request for SSD

In reply to Shaohua Li

For SSD, if request size exceeds specific value (optimal io size), request size isn't important for bandwidth. In such condition, if making request size bigger will cause some disks idle, the total throughput will actually drop. A good example is doing a readahead in a two-disk raid1 setup.

So when we should split big request? We absolutly don't want to split big request to very small requests. Even in SSD, big request transfer is more efficient. Below patch only consider request with size above optimal io size.

If all disks are busy, is it worthy to do split? Say optimal io size is 16k, two requests 32k and two disks. We can let each disk run one 32k request, or split the requests to 4 16k requests and each disk runs two. It's hard to say which case is better, depending on hardware.

So only consider case where there are idle disks. For readahead, split is always better in this case. And in my test, below patch can improve > 30% thoughput. Hmm, not 100%, because disk isn't 100% busy.

Such case can happen not just in readahead, for example, in directio. But I suppose directio usually will have bigger IO depth and make all disks busy, so I ignored it.

Signed-off-by Shaohua Li <[email protected]>
---
 drivers/md/raid1.c |   44 +++++++++++++++++++++++++++++++++++++-------
 drivers/md/raid1.h |    2 ++
 2 files changed, 39 insertions(+), 7 deletions(-)
drivers/md/raid1.c +37 -7
Index: linux/drivers/md/raid1.c
===================================================================
@@ -464,31 +464,49 @@ static void raid1_end_write_request(stru
 }
 
 static int read_balance_measure_ssd(struct r1conf *conf, struct r1bio *r1_bio,
-	int disk, int *best_disk, unsigned int *min_pending)
+	int disk, int *best_disk, unsigned int *min_pending, int *choose_idle)
 {
 	const sector_t this_sector = r1_bio->sector;
 	struct md_rdev *rdev;
 	unsigned int pending;
+	struct mirror_info *mirror = &conf->mirrors[disk];
+	int ret = 0;
 
-	rdev = rcu_dereference(conf->mirrors[disk].rdev);
+	rdev = rcu_dereference(mirror->rdev);
 	pending = atomic_read(&rdev->nr_pending);
 
 	/* big request IO helps SSD too, allow sequential IO merge */
-	if (conf->mirrors[disk].next_seq_sect == this_sector) {
+	if (mirror->next_seq_sect == this_sector && *choose_idle == 0) {
 		sector_t dist;
-		dist = abs(this_sector - conf->mirrors[disk].head_position);
+		dist = abs(this_sector - mirror->head_position);
 		/*
 		 * head_position is for finished request, such reqeust can't be
 		 * merged with current request, so it means nothing for SSD
 		 */
-		if (dist != 0)
+		if (dist != 0) {
+			/*
+			 * If buffered sequential IO size exceeds optimal
+			 * iosize and there is idle disk, choose idle disk
+			 */
+			if (mirror->seq_start != MaxSector
+			    && conf->opt_iosize > 0
+			    && mirror->next_seq_sect > conf->opt_iosize
+			    && mirror->next_seq_sect - conf->opt_iosize >=
+			       mirror->seq_start) {
+				*choose_idle = 1;
+				ret = 1;
+			}
 			goto done;
+		}
 	}
 
 	/* If device is idle, use it */
 	if (pending == 0)
 		goto done;
 
+	if (*choose_idle == 1)
+		return 1;
+
 	/* find device with less requests pending */
 	if (*min_pending > pending) {
 		*min_pending = pending;

@@ -497,7 +515,7 @@ static int read_balance_measure_ssd(stru
 	return 1;
 done:
 	*best_disk = disk;
-	return 0;
+	return ret;
 }
 
 static int read_balance_measure_distance(struct r1conf *conf,

@@ -551,6 +569,7 @@ static int read_balance(struct r1conf *c
 	unsigned int min_pending;
 	struct md_rdev *rdev;
 	int choose_first;
+	int choose_idle;
 
 	rcu_read_lock();
 	/*

@@ -564,6 +583,7 @@ static int read_balance(struct r1conf *c
 	best_dist = MaxSector;
 	min_pending = -1;
 	best_good_sectors = 0;
+	choose_idle = 0;
 
 	if (conf->mddev->recovery_cp < MaxSector &&
 	    (this_sector + sectors >= conf->next_resync))

@@ -647,7 +667,7 @@ static int read_balance(struct r1conf *c
 				break;
 		} else {
 			if (!read_balance_measure_ssd(conf, r1_bio, disk,
-			    &best_disk, &min_pending))
+			    &best_disk, &min_pending, &choose_idle))
 				break;
 		}
 	}

@@ -665,6 +685,10 @@ static int read_balance(struct r1conf *c
 			goto retry;
 		}
 		sectors = best_good_sectors;
+
+		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
+			conf->mirrors[best_disk].seq_start = this_sector;
+
 		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 	}
 	rcu_read_unlock();

@@ -2579,6 +2603,7 @@ static struct r1conf *setup_conf(struct
 	struct md_rdev *rdev;
 	int err = -ENOMEM;
 	bool nonrotational = true;
+	int opt_iosize = 0;
 
 	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
 	if (!conf)

@@ -2629,8 +2654,13 @@ static struct r1conf *setup_conf(struct
 		disk->head_position = 0;
 		if (!blk_queue_nonrot(bdev_get_queue(rdev->bdev)))
 			nonrotational = false;
+		else
+			opt_iosize = max(opt_iosize, bdev_io_opt(rdev->bdev));
+		disk->seq_start = MaxSector;
 	}
 	conf->nonrotational = nonrotational;
+	if (nonrotational)
+		conf->opt_iosize = opt_iosize >> 9;
 	conf->raid_disks = mddev->raid_disks;
 	conf->mddev = mddev;
 	INIT_LIST_HEAD(&conf->retry_list);
drivers/md/raid1.h +2 -0
Index: linux/drivers/md/raid1.h
===================================================================
@@ -9,6 +9,7 @@ struct mirror_info {
 	 * we try to keep sequential reads one the same device
 	 */
 	sector_t	next_seq_sect;
+	sector_t	seq_start;
 };
 
 /*

@@ -66,6 +67,7 @@ struct r1conf {
 	int			barrier;
 
 	int			nonrotational;
+	sector_t		opt_iosize;
 
 	/* Set to 1 if a full sync is needed, (fresh device added).
 	 * Cleared when a sync completes.
#5 lore raw
In reply to Shaohua Li
Quoted 8 lines · On Wed, 13 Jun 2012 17:09:22 +0800 Shaohua Li <[email protected]> wrote:

raid1 read balance is an important algorithm to make read performance optimal. It's a distance based algorithm, eg, for each request dispatch, choose disk whose last finished request is close the request. This is great for hard disk. But SSD has some special characteristics:

1. nonrotational. Distance means nothing for SSD, though merging small rquests to big request is still optimal for SSD. If no merge, distributing rquests to raid disks as more as possible is more optimal.

This, of course, has nothing to do with the devices rotating, and everything to do with there being a seek-penalty. So why the block-device flag is called "rotational" I really don't know :-(

Can we make the flags that md uses be "seek_penalty" or "no_seek" or something, even if the block layer has less meaningful names?

Quoted 7 lines

2. Getting too big request isn't always optimal. For hard disk, compared to spindle move, data transfer overhead is trival, so we always prefer bigger request. In SSD, request size exceeds specific value, performance isn't always increased with request size increased. An example is readahead. If readahead merges too big request and causes some disks idle, the performance is less optimal than that when all disks are busy and running small requests.

I note that the patch doesn't actually split requests, rather it avoids allocating adjacent requests to the same device - is that right? That seems sensible but doesn't seem to match your description so I wanted to check.

Quoted 4 lines

The patches try to address the issues. The first two patches are clean up. The third patch addresses the first item above. The forth addresses the second item above. The idea can be applied to raid10 too, which is in my todo list.

Thanks for these - there is a lot to like here.

One concern that I have has that they assume that all devices are the same. i.e. they are all "rotational", or none of them are.
they all have the same optimal IO size.

md aims to work well on heterogeneous arrays so I'll like to make it more general if I can.
Whether to "split" adjacent requests or not is decided in the context of a single device (the device that the first request is queued for) so that should be able to use the optimal IO size of that devices.

General balancing is a little harder as the decision is made in the context of all active devices. In particular we need to know how to choose between a seek-penalty device and a no-seek-penalty device, if they both have requests queued to them and the seek-penalty device is a long way from the target.

Maybe:
- if the last request to some device is within optimal-io-size of this requests, then send this request to that device.
- if either of two possible devices has no seek penalty, choose the one with the fewest outstanding requests.
- if both of two possible devices have a seek-penalty, then choose the closest

I think this will do the same as the current code for 'rotational' devices, and will be close to what your code does for 'non-rotational' devices.

There may well be room to improve this, but a key point is that it won't work to have two separate balancing routines - one for disks and one for ssds. I'm certainly happy for raid1_read_balance to be split up if it is too big, but I don't think that they way the first patch splits it is useful.

Would you be able to try something like that instead?

BTW, in a couple of places you use an 'rdev' taken out of the 'conf' structure without testing for NULL first. That isn't safe.
We get the 'rdev' at the top of the loop and test for NULL and other things. After that you need to always use *that* rdev value, don't try to get it out of the conf->mirrors structure again.

Thanks,
NeilBrown

#6 lore raw

hi guys,
some time ago i made some tests in raid1 md code, and a 'good' algorithm could improve 1% (mean) with some small variation (sometimes 3% sometime 0%) of performance boost (using new algorithm instead disk distance), in other words... read of 1GB could be improved to 1,01GB (10MB of improvement), well that was the result (about 1 year ago) and for a busy database server, that´s a good improvement (some guys tell that 0.01% of improvement is a improvement and should be used... well doesn´t matter...)

IMHO, i think that a general algorithm to select 'best device to read', should work with device information, and device information functions (sata/scsi/etc), i don´t know if it´s implemented but if not, should be implemented to work with this ideas....

putting code to select best distance is nice too, but it´s not general (think about a disk with 2 heads... a hardware raid1 for example...) the solution should be, get expected disks times to do some read job, and select the disk with smallest time
for example, instead putting a lot of statistics code in md, put this statistic in sata, scsi, or other 'under' device (nbd drbd ...) and their elevators functions (noop cfq, etc...)

check that md should implement a prediction function too, for example we could put a 2 pairs of raid1, and one raid0 over it, (a raid 1 + 0) and raid0 will contact raid1 devices to know what´s the best disk to read... well in this case just call 'unders' devices function to get a estimate time and select the 'best' disk

this is a bit 'hard' and a lot of work should be done, but IMHO, that´s what i think about 'ideal general solution'

at 'unders devices' (sata/scsi/drbd/nbd), today they could use statistic based on read/write times without big problems (this is good to have low cpu usage when we don´t know what´s the device is constructed), or maybe in future this could be implement at disk controller?! why not? see ssd drives, they implement a garbage collection! why not implement a 'time prediction function'?!

just some ideas....=)
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

author
#7 lore raw
In reply to NeilBrown
Quoted 11 lines · NeilBrown
Quoted 8 lines · On Wed, 13 Jun 2012 17:09:22 +0800 Shaohua Li <[email protected]> wrote:

raid1 read balance is an important algorithm to make read performance optimal. It's a distance based algorithm, eg, for each request dispatch, choose disk whose last finished request is close the request. This is great for hard disk. But SSD has some special characteristics:

1. nonrotational. Distance means nothing for SSD, though merging small rquests to big request is still optimal for SSD. If no merge, distributing rquests to raid disks as more as possible is more optimal.

Quoted 6 lines

This, of course, has nothing to do with the devices rotating, and everything to do with there being a seek-penalty. So why the block-device flag is called "rotational" I really don't know :-(

Can we make the flags that md uses be "seek_penalty" or "no_seek" or something, even if the block layer has less meaningful names?

purely borrowed from block layer. I'm ok to change it.

Quoted 8 lines
Quoted 7 lines

2. Getting too big request isn't always optimal. For hard disk, compared to spindle move, data transfer overhead is trival, so we always prefer bigger request. In SSD, request size exceeds specific value, performance isn't always increased with request size increased. An example is readahead. If readahead merges too big request and causes some disks idle, the performance is less optimal than that when all disks are busy and running small requests.

Quoted 4 lines

I note that the patch doesn't actually split requests, rather it avoids allocating adjacent requests to the same device - is that right? That seems sensible but doesn't seem to match your description so I wanted to check.

Yes, it doesn't do split, sorry for the misleading. I used to think about split, but gave up later, but forgot to change the description. There are two types of IO driving big request size: buffered read (readahead) buffered write and direct read or write I assume buffered write and direct read/write can drive high iodepth, so doing split is useless. buffered read split makes sense, which really should be 'prevent merging to big request'

Quoted 30 lines
Quoted 4 lines

The patches try to address the issues. The first two patches are clean up. The third patch addresses the first item above. The forth addresses the second item above. The idea can be applied to raid10 too, which is in my todo list.

Thanks for these - there is a lot to like here.

One concern that I have has that they assume that all devices are the same. i.e. they are all "rotational", or none of them are.
they all have the same optimal IO size.

md aims to work well on heterogeneous arrays so I'll like to make it more general if I can.
Whether to "split" adjacent requests or not is decided in the context of a single device (the device that the first request is queued for) so that should be able to use the optimal IO size of that devices.

General balancing is a little harder as the decision is made in the context of all active devices. In particular we need to know how to choose between a seek-penalty device and a no-seek-penalty device, if they both have requests queued to them and the seek-penalty device is a long way from the target.

Maybe:
- if the last request to some device is within optimal-io-size of this requests, then send this request to that device.
- if either of two possible devices has no seek penalty, choose the one with the fewest outstanding requests.
- if both of two possible devices have a seek-penalty, then choose the closest

Quoted 2 lines

I think this will do the same as the current code for 'rotational' devices, and will be close to what your code does for 'non-rotational' devices.

This is close to what I did except for the case of one hard disk and one SSD. Talking about heterogeneous arrary, I assume people only do it with two different hard disks or two different ssd. Will people really mix hard disk and SSD? A hard disk can only drive 600 IOPS while a lowend SATA SSD can drive 20k ~ 40k IOPS. Plusing 600 to 20k doesn't significantly change IOPS.

Quoted 5 lines

There may well be room to improve this, but a key point is that it won't work to have two separate balancing routines - one for disks and one for ssds. I'm certainly happy for raid1_read_balance to be split up if it is too big, but I don't think that they way the first patch splits it is useful. Would you be able to try something like that instead?

Yep, it's not worthy a separate function. My point is just making them independent. I can try this but not be convinced we need handle mixed harddisk/ssd case.

Quoted 5 lines

BTW, in a couple of places you use an 'rdev' taken out of the 'conf' structure without testing for NULL first. That isn't safe.
We get the 'rdev' at the top of the loop and test for NULL and other things. After that you need to always use *that* rdev value, don't try to get it out of the conf->mirrors structure again.

Thanks for pointing out, I'll fix it later.

Thanks,
Shaohua

#8 lore raw
Quoted 71 lines · On Thu, 28 Jun 2012 11:29:01 +0800 Shaohua Li <[email protected]> wrote:
Quoted 11 lines · NeilBrown
Quoted 8 lines · On Wed, 13 Jun 2012 17:09:22 +0800 Shaohua Li <[email protected]> wrote:

raid1 read balance is an important algorithm to make read performance optimal. It's a distance based algorithm, eg, for each request dispatch, choose disk whose last finished request is close the request. This is great for hard disk. But SSD has some special characteristics:

1. nonrotational. Distance means nothing for SSD, though merging small rquests to big request is still optimal for SSD. If no merge, distributing rquests to raid disks as more as possible is more optimal.

Quoted 6 lines

This, of course, has nothing to do with the devices rotating, and everything to do with there being a seek-penalty. So why the block-device flag is called "rotational" I really don't know :-(

Can we make the flags that md uses be "seek_penalty" or "no_seek" or something, even if the block layer has less meaningful names?

purely borrowed from block layer. I'm ok to change it.

Quoted 8 lines
Quoted 7 lines

2. Getting too big request isn't always optimal. For hard disk, compared to spindle move, data transfer overhead is trival, so we always prefer bigger request. In SSD, request size exceeds specific value, performance isn't always increased with request size increased. An example is readahead. If readahead merges too big request and causes some disks idle, the performance is less optimal than that when all disks are busy and running small requests.

Quoted 4 lines

I note that the patch doesn't actually split requests, rather it avoids allocating adjacent requests to the same device - is that right? That seems sensible but doesn't seem to match your description so I wanted to check.

Yes, it doesn't do split, sorry for the misleading. I used to think about split, but gave up later, but forgot to change the description. There are two types of IO driving big request size: buffered read (readahead) buffered write and direct read or write I assume buffered write and direct read/write can drive high iodepth, so doing split is useless. buffered read split makes sense, which really should be 'prevent merging to big request'

Quoted 30 lines
Quoted 4 lines

The patches try to address the issues. The first two patches are clean up. The third patch addresses the first item above. The forth addresses the second item above. The idea can be applied to raid10 too, which is in my todo list.

Thanks for these - there is a lot to like here.

One concern that I have has that they assume that all devices are the same. i.e. they are all "rotational", or none of them are.
they all have the same optimal IO size.

md aims to work well on heterogeneous arrays so I'll like to make it more general if I can.
Whether to "split" adjacent requests or not is decided in the context of a single device (the device that the first request is queued for) so that should be able to use the optimal IO size of that devices.

General balancing is a little harder as the decision is made in the context of all active devices. In particular we need to know how to choose between a seek-penalty device and a no-seek-penalty device, if they both have requests queued to them and the seek-penalty device is a long way from the target.

Maybe:
- if the last request to some device is within optimal-io-size of this requests, then send this request to that device.
- if either of two possible devices has no seek penalty, choose the one with the fewest outstanding requests.
- if both of two possible devices have a seek-penalty, then choose the closest

Quoted 8 lines
Quoted 2 lines

I think this will do the same as the current code for 'rotational' devices, and will be close to what your code does for 'non-rotational' devices.

This is close to what I did except for the case of one hard disk and one SSD. Talking about heterogeneous arrary, I assume people only do it with two different hard disks or two different ssd. Will people really mix hard disk and SSD? A hard disk can only drive 600 IOPS while a lowend SATA SSD can drive 20k ~ 40k IOPS. Plusing 600 to 20k doesn't significantly change IOPS.

I'm fairly sure people will try it, even if it doesn't make sense. :-)

It possibly would make sense if the slower device was marked "write-mostly", but that disables read-balance, so what we do with read balance might be irrelevant.

I don't want to assume devices have different speeds unless explicitly told so, and the "rotational" flag doesn't clearly imply anything about the speed of the device, only about seek latency. So e.g. I wouldn't want to always assume that a non-rotational device were faster than a rotational device, even though that may be the case in lots of situations.

I'd be less concerned about heterogeneous arrays with RAID10 as it makes less sense there. But I certainly want to at least acknowledge the possibility for RAID1.

thanks,
NeilBrown

#9 lore raw
Quoted 4 lines · NeilBrown
Quoted 2 lines · On Thu, 28 Jun 2012 11:29:01 +0800 Shaohua Li<[email protected]> wrote:
Quoted 1 line · NeilBrown

On Wed, 13 Jun 2012 17:09:22 +0800 Shaohua Li<[email protected]> wrote:

<snip>

Quoted 22 lines
Quoted 13 lines
Quoted 13 lines

General balancing is a little harder as the decision is made in the context of all active devices. In particular we need to know how to choose between a seek-penalty device and a no-seek-penalty device, if they both have requests queued to them and the seek-penalty device is a long way from the target.

Maybe:
- if the last request to some device is within optimal-io-size of this requests, then send this request to that device.
- if either of two possible devices has no seek penalty, choose the one with the fewest outstanding requests.
- if both of two possible devices have a seek-penalty, then choose the closest

Quoted 8 lines
Quoted 2 lines

I think this will do the same as the current code for 'rotational' devices, and will be close to what your code does for 'non-rotational' devices.

This is close to what I did except for the case of one hard disk and one SSD. Talking about heterogeneous arrary, I assume people only do it with two different hard disks or two different ssd. Will people really mix hard disk and SSD? A hard disk can only drive 600 IOPS while a lowend SATA SSD can drive 20k ~ 40k IOPS. Plusing 600 to 20k doesn't significantly change IOPS.

Quoted 5 lines

I'm fairly sure people will try it, even if it doesn't make sense. :-)

It possibly would make sense if the slower device was marked "write-mostly", but that disables read-balance, so what we do with read balance might be irrelevant.

I would imagine that an SSD paired with an HD (or a partition on an HD) in RAID1 would be very common for small setups - at least until the current experimental flash cache systems solidify and mature. You get the speed of the SSD, and the safety of RAID1 without the cost of two SSDs. It is normal today to use write-mostly on the HD - but I'd guess a fair number of users don't think to do that. And with good balancing from these patches, there is no need - and you can get the best from both devices. SSD's are much faster than HD's when there is seeking involved - but for streamed accesses the difference is smaller, and it's good to have both devices available for reads.

Quoted 14 lines

I don't want to assume devices have different speeds unless explicitly told so, and the "rotational" flag doesn't clearly imply anything about the speed of the device, only about seek latency. So e.g. I wouldn't want to always assume that a non-rotational device were faster than a rotational device, even though that may be the case in lots of situations.

I'd be less concerned about heterogeneous arrays with RAID10 as it makes less sense there. But I certainly want to at least acknowledge the possibility for RAID1.

thanks,
NeilBrown