Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Already on GitHub? Sign in to your account

Improve N-way mirror performance #1487

Closed
wants to merge 1 commit into
from

Conversation

Projects
None yet
3 participants
Owner

behlendorf commented May 31, 2013

The read bandwidth of an N-way mirror can by increased by 50%,
and the IOPs by 10%, by more carefully selecting the preferred
leaf vdev.

The existing algorthm selects a perferred leaf vdev based on
offset of the zio request modulo the number of members in the
mirror. It assumes the drives are of equal performance and
that spreading the requests randomly over both drives will be
sufficient to saturate them. In practice this results in the
leaf vdevs being under utilized.

Utilization can be improved by preferentially selecting the leaf
vdev with the least pending IO. This prevents leaf vdevs from
being starved and compensates for performance differences between
disks in the mirror. Faster vdevs will be sent more work and
the mirror performance will not be limitted by the slowest drive.

In the common case where all the pending queues are full and there
is no single least busy leaf vdev a batching stratagy is employed.
Of the N least busy vdevs one is selected with equal probability
to be the preferred vdev for T milliseconds. Compared to randomly
selecting a vdev to break the tie batching the requests greatly
improves the odds of merging the requests in the Linux elevator.

The testing results show a significant performance improvement
for all four workloads tested. The workloads were generated
using the fio benchmark and are as follows.

1) 1MB sequential reads from 16 threads to 16 files (MB/s).
2) 4KB sequential reads from 16 threads to 16 files (MB/s).
3) 1MB random reads from 16 threads to 16 files (IOP/s).
4) 4KB random reads from 16 threads to 16 files (IOP/s).

               | Pristine              |  With 1461             |
               | Sequential  Random    |  Sequential  Random    |
               | 1MB  4KB    1MB  4KB  |  1MB  4KB    1MB  4KB  |
               | MB/s MB/s   IO/s IO/s |  MB/s MB/s   IO/s IO/s |
---------------+-----------------------+------------------------+
2 Striped      | 226  243     11  304  |  222  255     11  299  |
2 2-Way Mirror | 302  324     16  534  |  433  448     23  571  |
2 3-Way Mirror | 429  458     24  714  |  648  648     41  808  |
2 4-Way Mirror | 562  601     36  849  |  816  828     82  926  |

Signed-off-by: Brian Behlendorf behlendorf1@llnl.gov
Issue #1461

@behlendorf behlendorf Improve N-way mirror performance
The read bandwidth of an N-way mirror can by increased by 50%,
and the IOPs by 10%, by more carefully selecting the preferred
leaf vdev.

The existing algorthm selects a perferred leaf vdev based on
offset of the zio request modulo the number of members in the
mirror.  It assumes the drives are of equal performance and
that spreading the requests randomly over both drives will be
sufficient to saturate them.  In practice this results in the
leaf vdevs being under utilized.

Utilization can be improved by preferentially selecting the leaf
vdev with the least pending IO.  This prevents leaf vdevs from
being starved and compensates for performance differences between
disks in the mirror.  Faster vdevs will be sent more work and
the mirror performance will not be limitted by the slowest drive.

In the common case where all the pending queues are full and there
is no single least busy leaf vdev a batching stratagy is employed.
Of the N least busy vdevs one is selected with equal probability
to be the preferred vdev for T milliseconds.  Compared to randomly
selecting a vdev to break the tie batching the requests greatly
improves the odds of merging the requests in the Linux elevator.

The testing results show a significant performance improvement
for all four workloads tested.  The workloads were generated
using the fio benchmark and are as follows.

1) 1MB sequential reads from 16 threads to 16 files (MB/s).
2) 4KB sequential reads from 16 threads to 16 files (MB/s).
3) 1MB random reads from 16 threads to 16 files (IOP/s).
4) 4KB random reads from 16 threads to 16 files (IOP/s).

               | Pristine              |  With 1461             |
               | Sequential  Random    |  Sequential  Random    |
               | 1MB  4KB    1MB  4KB  |  1MB  4KB    1MB  4KB  |
               | MB/s MB/s   IO/s IO/s |  MB/s MB/s   IO/s IO/s |
---------------+-----------------------+------------------------+
2 Striped      | 226  243     11  304  |  222  255     11  299  |
2 2-Way Mirror | 302  324     16  534  |  433  448     23  571  |
2 3-Way Mirror | 429  458     24  714  |  648  648     41  808  |
2 4-Way Mirror | 562  601     36  849  |  816  828     82  926  |

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1461
20ad617
Contributor

pyavdr commented Jun 2, 2013

@behlendorf
Hi Brian,

i use a SUSE zfs server using a zpool with 4 striped two way mirrors to check that patch.
ZFS is serving zvols ( sparse, no comp, no sync ) via scst to Windows. Windows formats a ntfs file system on that iscsi share (10 Gbit/s). With CrystalDiskMark 3.0.1 i got the following performance
values:

with your patch:

       Sequential Read :   558.050 MB/s
      Sequential Write :   330.000 MB/s
     Random Read 512KB :   470.839 MB/s
    Random Write 512KB :   225.294 MB/s
Random Read 4KB (QD=1) :    19.295 MB/s [  4710.6 IOPS]

Random Write 4KB (QD=1) : 11.762 MB/s [ 2871.6 IOPS]
Random Read 4KB (QD=32) : 168.782 MB/s [ 41206.5 IOPS]
Random Write 4KB (QD=32) : 10.404 MB/s [ 2540.0 IOPS]

without your patch:

      Sequential Read :   526.900 MB/s
      Sequential Write :   394.600 MB/s
     Random Read 512KB :   458.800 MB/s
    Random Write 512KB :   374.6 MB/s
Random Read 4KB (QD=1) :    16.780 MB/s 

Random Write 4KB (QD=1) : 15.670 MB/s
Random Read 4KB (QD=32) : 224.500 MB/s
Random Write 4KB (QD=32) : 25.410 MB/s

The patch clearly improves seq. read performance but random
access is far better without that patch. It might work better
on 3 or 4 way mirrors, but not on my striped 2 way mirrors.

Contributor

GregorKopka commented Jun 3, 2013

@pyavdr: What settings do you have on the ZVOL (zfs get all)? Has dedup ever been enabled on the pool?
Since the patch won't touch the write path i'm curious where the missing 150MB/s random writes come from.

Also: maybe you could try b333z/zfs@ac0df7e (the new code needs to be activated with echo 1 > /sys/module/zfs/parameters/zfs_vdev_mirror_pending_balance) which has a different approach?

Contributor

pyavdr commented Jun 3, 2013

@GregorKopka

no dedup at all. zfs settings:

NAME PROPERTY VALUE SOURCE
stor1/ntfssh type volume -
stor1/ntfssh creation Sun Jun 2 16:36 2013 -
stor1/ntfssh used 7.83G -
stor1/ntfssh available 4.30T -
stor1/ntfssh referenced 7.83G -
stor1/ntfssh compressratio 1.00x -
stor1/ntfssh reservation none default
stor1/ntfssh volsize 1000G local
stor1/ntfssh volblocksize 64K -
stor1/ntfssh checksum on default
stor1/ntfssh compression off local
stor1/ntfssh readonly off default
stor1/ntfssh copies 1 default
stor1/ntfssh refreservation none default
stor1/ntfssh primarycache all default
stor1/ntfssh secondarycache all default
stor1/ntfssh usedbysnapshots 0 -
stor1/ntfssh usedbydataset 7.83G -
stor1/ntfssh usedbychildren 0 -
stor1/ntfssh usedbyrefreservation 0 -
stor1/ntfssh logbias latency default
stor1/ntfssh dedup off default
stor1/ntfssh mlslabel none default
stor1/ntfssh sync disabled local
stor1/ntfssh refcompressratio 1.00x -
stor1/ntfssh written 7.83G -
stor1/ntfssh snapdev hidden default

Contributor

pyavdr commented Jun 4, 2013

@GregorKopka
I would say the patch b333z@ac0df7e decreases performance in my case.

without patch

       Sequential Read :   488.107 MB/s
      Sequential Write :   395.801 MB/s
     Random Read 512KB :   454.163 MB/s
    Random Write 512KB :   345.225 MB/s
Random Read 4KB (QD=1) :    17.392 MB/s [  4246.1 IOPS]

Random Write 4KB (QD=1) : 16.281 MB/s [ 3974.8 IOPS]
Random Read 4KB (QD=32) : 219.449 MB/s [ 53576.5 IOPS]
Random Write 4KB (QD=32) : 23.440 MB/s [ 5722.6 IOPS]

with patch b333z@ac0df7e

       Sequential Read :   485.340 MB/s
      Sequential Write :   323.435 MB/s
     Random Read 512KB :   421.094 MB/s
    Random Write 512KB :   176.305 MB/s
Random Read 4KB (QD=1) :    16.779 MB/s [  4096.4 IOPS]

Random Write 4KB (QD=1) : 15.526 MB/s [ 3790.6 IOPS]
Random Read 4KB (QD=32) : 149.005 MB/s [ 36378.3 IOPS]
Random Write 4KB (QD=32) : 20.374 MB/s [ 4974.2 IOPS]

Owner

behlendorf commented Jul 11, 2013

After changing the zfs_vdev_mirror_switch units from ms to us this was merged as:

556011d Improve N-way mirror performance

@behlendorf behlendorf closed this Jul 11, 2013

@b333z b333z added a commit to b333z/zfs that referenced this pull request Oct 30, 2013

@b333z b333z Improve ZFS N-way mirror read performance + Load/Locality/Rotation
Freebsd Rev. 256956 by @stevenh - Improve ZFS N-way mirror read performance by using load and locality
information.

Reviewed by: gibbs, mav, will
Sponsored by: Multiplay

References:
  http://svnweb.freebsd.org/changeset/base/256956
  zfsonlinux#1487
  zfsonlinux#1803
  http://open-zfs.org/wiki/Features#Improve_N-way_mirror_read_performance

Ported by: Andrew Barnes <barnes333@gmail.com>

Closes #1803
bbe360b

@heliocentric heliocentric pushed a commit to heliocentric/freebsd that referenced this pull request Oct 31, 2013

smh Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay
5c7a6f5

@heliocentric heliocentric pushed a commit to heliocentric/freebsd that referenced this pull request Oct 31, 2013

smh Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay


git-svn-id: svn+ssh://svn.freebsd.org/base/head@256956 ccf9f872-aa2e-dd11-9fc8-001c23d0bc1f
25f0cfe

@delphij delphij added a commit to freenas/os that referenced this pull request Feb 5, 2014

@delphij smh + delphij Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

(cherry picked from commit 5c7a6f5)
f0962d7

@ixgitserver ixgitserver pushed a commit to freenas/os that referenced this pull request May 27, 2014

smh + Craig Rodrigues Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

(cherry picked from commit 5c7a6f5)
ab45b41

@dspinellis dspinellis pushed a commit to dspinellis/unix-history-repo that referenced this pull request Jul 29, 2014

Steven Hartland Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay


git-svn-id: svn+ssh://svn.freebsd.org/base/head@256956 ccf9f872-aa2e-dd11-9fc8-001c23d0bc1f
0d6087d

@delphij delphij added a commit to freenas/os that referenced this pull request Aug 5, 2014

@delphij smh + delphij Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

(cherry picked from commit 5c7a6f5)
3ffa592

@delphij delphij added a commit to freenas/os that referenced this pull request Aug 7, 2014

@delphij smh + delphij Improve ZFS N-way mirror read performance by using load and locality
information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

(cherry picked from commit 5c7a6f5)
25dcbd5

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 13, 2016

@ryao smh + ryao Improve ZFS N-way mirror read performance by using load and locality …
…information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

Ported-by: Richard Yao <ryao@gentoo.org>
0fb4a64

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 13, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

Ported-by: Richard Yao <ryao@gentoo.org>
a6a040c

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 13, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

Ported-by: Richard Yao <ryao@gentoo.org>
48c4cde

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 16, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

Ported-by: Richard Yao <ryao@gentoo.org>
fb1e93e

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 24, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
68251c7

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 24, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
8e18498

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 24, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- vdev_mirror_shift from OpenSolaris was missing from our code, so it
  has been added.

- A typo "IO'a" has been corrected to say "IO's"

- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
1bffa12

@ryao ryao added a commit to ryao/zfs that referenced this pull request Feb 24, 2016

@ryao smh + ryao FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

Porting notes:

- The tunables were adjusted to have ZoL-style names.

- The code was modified to use ZoL's vd_nonrot.

- Fixes were done to make cstyle.pl happy

- Merge conflicts were handled manually

- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.

- This replaces 556011d entirely.

- A typo "IO'a" has been corrected to say "IO's"

- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
f19d289

@behlendorf behlendorf added a commit that referenced this pull request Feb 26, 2016

@behlendorf smh + behlendorf FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

References:
  https://github.com/freebsd/freebsd@5c7a6f5d
  https://github.com/freebsd/freebsd@31b7f68d
  https://github.com/freebsd/freebsd@e186f564

Performance Testing:
  #4334 (comment)

Porting notes:
- The tunables were adjusted to have ZoL-style names.
- The code was modified to use ZoL's vd_nonrot.
- Fixes were done to make cstyle.pl happy
- Merge conflicts were handled manually
- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.
- This replaces 556011d entirely.
- A typo "IO'a" has been corrected to say "IO's"
- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #4334
9f50093

@lundman lundman added a commit to openzfsonosx/zfs that referenced this pull request Feb 29, 2016

@lundman smh + lundman FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

References:
  https://github.com/freebsd/freebsd@5c7a6f5d
  https://github.com/freebsd/freebsd@31b7f68d
  https://github.com/freebsd/freebsd@e186f564

Performance Testing:
  zfsonlinux/zfs#4334 (comment)

Porting notes:
- The tunables were adjusted to have ZoL-style names.
- The code was modified to use ZoL's vd_nonrot.
- Fixes were done to make cstyle.pl happy
- Merge conflicts were handled manually
- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.
- This replaces 556011d entirely.
- A typo "IO'a" has been corrected to say "IO's"
- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

Changed kstat types, and added kstat defines for OSX.

Ported-by: Jorgen Lundman <lundman@lundman.net>
d746c3a

@rottegift rottegift pushed a commit to rottegift/zfs that referenced this pull request Mar 1, 2016

smh + rottegift FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

References:
  https://github.com/freebsd/freebsd@5c7a6f5d
  https://github.com/freebsd/freebsd@31b7f68d
  https://github.com/freebsd/freebsd@e186f564

Performance Testing:
  zfsonlinux/zfs#4334 (comment)

Porting notes:
- The tunables were adjusted to have ZoL-style names.
- The code was modified to use ZoL's vd_nonrot.
- Fixes were done to make cstyle.pl happy
- Merge conflicts were handled manually
- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.
- This replaces 556011d entirely.
- A typo "IO'a" has been corrected to say "IO's"
- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

Changed kstat types, and added kstat defines for OSX.

Ported-by: Jorgen Lundman <lundman@lundman.net>
dec7699

@lundman lundman added a commit to openzfsonosx/zfs that referenced this pull request Jan 23, 2017

@lundman smh + lundman FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

References:
  https://github.com/freebsd/freebsd@5c7a6f5d
  https://github.com/freebsd/freebsd@31b7f68d
  https://github.com/freebsd/freebsd@e186f564

Performance Testing:
  zfsonlinux/zfs#4334 (comment)

Porting notes:
- The tunables were adjusted to have ZoL-style names.
- The code was modified to use ZoL's vd_nonrot.
- Fixes were done to make cstyle.pl happy
- Merge conflicts were handled manually
- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.
- This replaces 556011d entirely.
- A typo "IO'a" has been corrected to say "IO's"
- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

Changed kstat types, and added kstat defines for OSX.

Ported-by: Jorgen Lundman <lundman@lundman.net>
ccdcf85

@lundman lundman added a commit to openzfsonosx/zfs that referenced this pull request Jan 24, 2017

@lundman smh + lundman FreeBSD r256956: Improve ZFS N-way mirror read performance by using l…
…oad and locality information.

The existing algorithm selects a preferred leaf vdev based on offset of the zio
request modulo the number of members in the mirror. It assumes the devices are
of equal performance and that spreading the requests randomly over both drives
will be sufficient to saturate them. In practice this results in the leaf vdevs
being under utilized.

The new algorithm takes into the following additional factors:
* Load of the vdevs (number outstanding I/O requests)
* The locality of last queued I/O vs the new I/O request.

Within the locality calculation additional knowledge about the underlying vdev
is considered such as; is the device backing the vdev a rotating media device.

This results in performance increases across the board as well as significant
increases for predominantly streaming loads and for configurations which don't
have evenly performing devices.

The following are results from a setup with 3 Way Mirror with 2 x HD's and
1 x SSD from a basic test running multiple parrallel dd's.

With pre-fetch disabled (vfs.zfs.prefetch_disable=1):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 161 seconds @ 95 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 297 seconds @ 51 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 54 seconds @ 284 MB/s

With pre-fetch enabled (vfs.zfs.prefetch_disable=0):

== Stripe Balanced (default) ==
Read 15360MB using bs: 1048576, readers: 3, took 91 seconds @ 168 MB/s
== Load Balanced (zfslinux) ==
Read 15360MB using bs: 1048576, readers: 3, took 108 seconds @ 142 MB/s
== Load Balanced (locality freebsd) ==
Read 15360MB using bs: 1048576, readers: 3, took 48 seconds @ 320 MB/s

In addition to the performance changes the code was also restructured, with
the help of Justin Gibbs, to provide a more logical flow which also ensures
vdevs loads are only calculated from the set of valid candidates.

The following additional sysctls where added to allow the administrator
to tune the behaviour of the load algorithm:
* vfs.zfs.vdev.mirror.rotating_inc
* vfs.zfs.vdev.mirror.rotating_seek_inc
* vfs.zfs.vdev.mirror.rotating_seek_offset
* vfs.zfs.vdev.mirror.non_rotating_inc
* vfs.zfs.vdev.mirror.non_rotating_seek_inc

These changes where based on work started by the zfsonlinux developers:
zfsonlinux/zfs#1487

Reviewed by:	gibbs, mav, will
MFC after:	2 weeks
Sponsored by:	Multiplay

References:
  https://github.com/freebsd/freebsd@5c7a6f5d
  https://github.com/freebsd/freebsd@31b7f68d
  https://github.com/freebsd/freebsd@e186f564

Performance Testing:
  zfsonlinux/zfs#4334 (comment)

Porting notes:
- The tunables were adjusted to have ZoL-style names.
- The code was modified to use ZoL's vd_nonrot.
- Fixes were done to make cstyle.pl happy
- Merge conflicts were handled manually
- freebsd/freebsd@e186f56 by my
  collegue Andriy Gapon has been included. It applied perfectly, but
  added a cstyle regression.
- This replaces 556011d entirely.
- A typo "IO'a" has been corrected to say "IO's"
- Descriptions of new tunables were added to man/man5/zfs-module-parameters.5.

Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

Changed kstat types, and added kstat defines for OSX.

Ported-by: Jorgen Lundman <lundman@lundman.net>
703154b

@behlendorf behlendorf deleted the behlendorf:issue-1461 branch Feb 16, 2017

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment