SSD Acceleration: Difference between revisions

From NURDspace
No edit summary
Line 250: Line 250:
* read, 162.114MB/s, 2.716ms, 20750 IOps
* read, 162.114MB/s, 2.716ms, 20750 IOps
* write, 35.676MB/s, 13.936ms, 4566 IOps
* write, 35.676MB/s, 13.936ms, 4566 IOps
and with writethrough ;
* randread, 160.338MB/s, 2.795ms, 20523 IOps
* randrw, 1.302MB/s, 226.778ms, 166 IOps
* read, 146.139MB/s, 3.089ms, 18705 IOps
* write, 7.118MB/s, 63.971ms, 911 IOps


Full results here ; [[SSD Acceleration/Toshiba MQ01ABD100 with bcache]]
Full results here ; [[SSD Acceleration/Toshiba MQ01ABD100 with bcache]]

Revision as of 01:44, 4 March 2017


SSD Acceleration test
Ssdnow.jpg
Participants
Skills Linux
Status Active
Niche Software
Purpose Infra
Tool No
Location Server room
Cost 0
Tool category

SSD Acceleration test

ssdnow.jpg {{#if:No | [[Tool Owner::{{{ProjectParticipants}}} | }} {{#if:No | [[Tool Cost::0 | }}

In preparation of building a new storage environment, I wanted to see if SSD acceleration would help improve IOPS and throughput on a classic HDD array.

attempts

attempt 1

In preparation for this, I put together the following machine: An ASUS P5S800-VM with:

  • 1Gb RAM
  • 1x Intel(R) Celeron(R) CPU 2.66GHz
  • 1x Hitachi Deskstar IC35L120AVV207 'boot' disk
    • IDE
    • 241254720 sectors @ 4096 bits/s (115 Gb)
  • 1x Hitachi HDT721010SLA360 'storage' disk
    • SATA
    • 1953525168 sectors @ 4096 bits/s (931.5 Gb)
  • 1x Kingston SV300S37A120G 'caching' SSD
    • SATA
    • 234441648 sectors @ 4096 bits/s (111.7 Gb)

Using fio and https://github.com/tsaikd/fio-wrapper , I have tested the raw MB/s and IOPS for each device, using the default 32Mb file size: (Graphs coming soon)

  • Boot:
    • randread, 64 users, 8K blocks: 1.7 MB/s , 288 ms latency, 219 IOPS
    • randrw, 64 users, 8K blocks: 0.68 MB/s , 385 ms latency, 86 IOPS
    • read, 64 users, 8K blocks: 32 MB/s , 15 ms latency, 4207 IOPS
    • write, 64 users, 8K blocks: 20 MB/s , 24 ms latency, 2569 IOPS
  • Storage:
    • randread, 64 users, 8K blocks: 1.7 MB/s , 283 ms latency, 222 IOPS
    • randrw, 64 users, 8K blocks: 0.832 MB/s , 305 ms latency, 106 IOPS
    • read, 64 users, 8K blocks: 34.2 MB/s , 14.5 ms latency, 4381 IOPS
    • write, 64 users, 8K blocks: 17.4 MB/s , 27 ms latency, 2220 IOPS
  • Caching:
    • randread, 64 users, 8K blocks: 8 MB/s , 0.46 ms latency, 2060 IOPS
    • randrw, 64 users, 8K blocks: 10 MB/s , 23.7 ms latency, 1348 IOPS
    • read, 64 users, 8K blocks: 29.7 MB/s , 16.2 ms latency, 3811 IOPS
    • write, 64 users, 8K blocks: 20.9 MB/s , 23.6 ms latency, 2683 IOPS

We can see the SSD provides about the same QOS for linear rw operations, but is significantly better at random rw than the spinning disks. This is not unexpected.

dmsetup cache

A write-back dmsetup cache system was constructed using the following configuration:

root@roger-wilco:~# dmsetup ls
roger--wilco--vg-swap   (254:2)
roger--wilco--vg-root   (254:0)
ssd-metadata    (254:3)
ssd-blocks      (254:4)
cached-disk     (254:6)
vg0-spindle     (254:5)

root@roger-wilco:~# dmsetup table ssd-metadata
0 924000 linear 8:33 0
root@roger-wilco:~# dmsetup table ssd-blocks
0 233515600 linear 8:33 924000
root@roger-wilco:~# dmsetup table cached-disk
0 1953513472 cache 254:3 254:4 254:5 512 1 writeback default 0

root@roger-wilco:~# vgs
 VG             #PV #LV #SN Attr   VSize   VFree
 roger-wilco-vg   1   3   0 wz--n- 114.80g    0
 vg0              1   1   0 wz--n- 931.51g    0
root@roger-wilco:~# pvs
 PV         VG             Fmt  Attr PSize   PFree
 /dev/sda5  roger-wilco-vg lvm2 a--  114.80g    0
 /dev/sdb1  vg0            lvm2 a--  931.51g    0
root@roger-wilco:~# lvs
 LV      VG             Attr       LSize   Pool Origin Data%  Meta%  Move Log Cpy%Sync Convert
 home    roger-wilco-vg -wi-ao---- 103.49g
 root    roger-wilco-vg -wi-ao----   9.31g
 swap    roger-wilco-vg -wi-ao----   2.00g
 spindle vg0            -wi-ao---- 931.51g
  • Cached:
    • randread, 64 users, 8K blocks: 26.7 MB/s , 18.5 ms latency, 3423 IOPS
    • randrw, 64 users, 8K blocks: 5.36 MB/s , 46.6 ms latency, 686 IOPS
    • read, 64 users, 8K blocks: 30.1 MB/s , 16.5 ms latency, 3857 IOPS
    • write, 64 users, 8K blocks: 1 MB/s , 477 ms latency, 127 IOPS

With this setup, we see a tradeoff between linear write and random write. Linear write is significantly slowed for some reason - some issue with the writeback configuration?

  • Cached (writethrough):
    • randread, 64 users, 8K blocks: 0.8 MB/s , 598 ms latency, 102 IOPS
    • randrw, 64 users, 8K blocks: 0.4 MB/s , 601 ms latency, 52 IOPS
    • read, 64 users, 8K blocks: 30 MB/s , 16.5 ms latency, 3836 IOPS
    • write, 64 users, 8K blocks: 9 MB/s , 53 ms latency, 1160 IOPS

NOPE

lvm cache

OK, so there's another means of building caching:

root@roger-wilco:~# vgextend vg0 /dev/sdc1
 Physical volume "/dev/sdc1" successfully created
 Volume group "vg0" successfully extended
root@roger-wilco:~# pvs
 PV         VG             Fmt  Attr PSize   PFree  
 /dev/sda5  roger-wilco-vg lvm2 a--  114.80g      0 
 /dev/sdb1  vg0            lvm2 a--  931.51g      0 
 /dev/sdc1  vg0            lvm2 a--  111.79g 111.79g
root@roger-wilco:~# lvs
 LV      VG             Attr       LSize   Pool Origin Data%  Meta%  Move Log Cpy%Sync Convert
 home    roger-wilco-vg -wi-ao---- 103.49g                                                    
 root    roger-wilco-vg -wi-ao----   9.31g                                                    
 swap    roger-wilco-vg -wi-ao----   2.00g                                                    
 spindle vg0            -wi-a----- 931.51g                                                    
root@roger-wilco:~# lvcreate -L 1G cache_meta vg0 /dev/sdc1
 Volume group "cache_meta" not found
root@roger-wilco:~# lvcreate -L 1G -n cache_meta vg0 /dev/sdc1
 Logical volume "cache_meta" created
root@roger-wilco:~# lvcreate -l 100%FREE -n cache vg0 /dev/sdc1
 Logical volume "cache" created
root@roger-wilco:~# lvs
 LV         VG             Attr       LSize   Pool Origin Data%  Meta%  Move Log Cpy%Sync Convert
 home       roger-wilco-vg -wi-ao---- 103.49g                                                    
 root       roger-wilco-vg -wi-ao----   9.31g                                                    
 swap       roger-wilco-vg -wi-ao----   2.00g                                                    
 cache      vg0            -wi-a----- 110.79g                                                    
 cache_meta vg0            -wi-a-----   1.00g                                                    
 spindle    vg0            -wi-a----- 931.51g                                                    
root@roger-wilco:~# lvconvert --type cache-pool --poolmetadata vg0/cache_meta vg0/cache
 WARNING: Converting logical volume vg0/cache and vg0/cache_meta to pool's data and metadata volumes.
 THIS WILL DESTROY CONTENT OF LOGICAL VOLUME (filesystem etc.)
Do you really want to convert vg0/cache and vg0/cache_meta? [y/n]: y
 Volume group "vg0" has insufficient free space (0 extents): 256 required.
root@roger-wilco:~# lvconvert --type cache vg0/cache vg0/spindle  --cache requires --cachepool.
 Run `lvconvert --help' for more information.
root@roger-wilco:~# lvconvert --type cache --cachepool vg0/cache vg0/spindle
 WARNING: Converting logical volume vg0/cache to pool's data volume.
 THIS WILL DESTROY CONTENT OF LOGICAL VOLUME (filesystem etc.)
Do you really want to convert vg0/cache? [y/n]: y
 Rounding up size to full physical extent 60.00 MiB
 Volume group "vg0" has insufficient free space (0 extents): 15 required.

(ER...)

root@roger-wilco:~# e2fsck -f /dev/mapper/vg0-spindle 
...
root@roger-wilco:~# resize2fs /dev/mapper/vg0-spindle 300G
...
root@roger-wilco:~# lvreduce -L 500G vg0/spindle
 WARNING: Reducing active logical volume to 500.00 GiB
 THIS MAY DESTROY YOUR DATA (filesystem etc.)
Do you really want to reduce spindle? [y/n]: y
 Size of logical volume vg0/spindle changed from 931.51 GiB (238466 extents) to 500.00 GiB (128000 extents).
 Logical volume spindle successfully resized
root@roger-wilco:~# lvconvert --type cache --cachepool vg0/cache vg0/spindle
 WARNING: Converting logical volume vg0/cache to pool's data volume.
 THIS WILL DESTROY CONTENT OF LOGICAL VOLUME (filesystem etc.)
Do you really want to convert vg0/cache? [y/n]: y
 Rounding up size to full physical extent 60.00 MiB
 Logical volume "lvol0" created
 Logical volume "lvol0" created
 Converted vg0/cache to cache pool.
 Logical volume vg0/spindle is now cached.
root@roger-wilco:~# lvresize -L +100%FREE vg0/spindle
 Invalid argument for --size: +100%FREE
 Error during parsing of command line.
root@roger-wilco:~# lvresize -l +100%FREE vg0/spindle
 Unable to resize logical volumes of cache type.

(ER.?!)

root@roger-wilco:~# lvremove vg0/spindle
Do you really want to remove active logical volume spindle? [y/n]: y
 /usr/sbin/cache_check: execvp failed: No such file or directory
 WARNING: Integrity check of metadata for pool vg0/cache failed.
 Logical volume "spindle" successfully removed
root@roger-wilco:~# vgs
 VG             #PV #LV #SN Attr   VSize   VFree  
 roger-wilco-vg   1   3   0 wz--n- 114.80g      0 
 vg0              2   2   0 wz--n-   1.02t 931.39g
root@roger-wilco:~# lvcreate -L 930G -n spindle vg0 /dev/sdb1
 Logical volume "spindle" created
root@roger-wilco:~# lvconvert --type cache --cachepool vg0/cache vg0/spindle
 Logical volume vg0/spindle is now cached.
root@roger-wilco:~# mkfs.ext4 /dev/mapper/vg0-spindle
mke2fs 1.42.12 (29-Aug-2014)
/dev/mapper/vg0-spindle contains a ext4 file system
  last mounted on /mnt/cached on Wed Apr  6 11:54:51 2016
Proceed anyway? (y,n) n

(OK...)

root@roger-wilco:~# mount /dev/mapper/vg0-spindle /mnt/cached/


attempt 2

For the oncoming downscaling of the server infra, we needed a new target. The question of SSD cache came up, but was discarded as not-for-production-yet. Never the less, i tested it's performance;

setup

  • Intel Pentium J4205 ( Asrock J4205-ITX ), 16GB 1600Mhz DDR3
  • HDD; 1TB Toshiba MQ01ABD100 5400rpm, 8MB cache
  • SDD; 120GB Kingston SUV400S37120G


Disk /dev/sda: 931.5 GiB, 1000204886016 bytes, 1953525168 sectors
Units: sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 4096 bytes
I/O size (minimum/optimal): 4096 bytes / 4096 bytes
Disklabel type: gpt
Disk identifier: 28FE43B1-093A-4350-8E02-E3383C59A994

Device         Start        End    Sectors   Size Type
/dev/sda1       2048       4095       2048     1M BIOS boot
/dev/sda2       4096     528383     524288   256M EFI System
/dev/sda3     528384  125528063  124999680  59.6G Linux filesystem
/dev/sda4  125528064  188028927   62500864  29.8G Linux swap
/dev/sda5  188028928 1953525134 1765496207 841.9G Linux filesystem


Only /dev/sda5 is used to test.

Plain disk results

(all results 64 users, 8K blocks for comparison)

  • randread, 2.909MB/s, 170.255ms, 372 IOps
  • randrw, 0.476MB/s, 657.810ms, 60 IOps
  • read, 36.373MB/s, 13.671ms, 4655 IOps
  • write, 7.488MB/s, 62.896ms, 958 IOps

Full results here ; SSD Acceleration/Toshiba MQ01ABD100

bcache results

I setup the full ssd (/dev/sdb) as cache device for the backingstore (/dev/sda5). I also enable writeback and discard by default.

apt install bcache-tools
make-bcache -C /dev/sdb -B /dev/sda5 --discard --writeback
mkfs.ext4 /dev/bcache0

(all results 64 users, 8K blocks for comparison)

  • randread, 177.577MB/s, 2.505ms, 22729 IOps
  • randrw, 10.483MB/s, 31.560ms, 1341 IOps
  • read, 162.114MB/s, 2.716ms, 20750 IOps
  • write, 35.676MB/s, 13.936ms, 4566 IOps

and with writethrough ;

  • randread, 160.338MB/s, 2.795ms, 20523 IOps
  • randrw, 1.302MB/s, 226.778ms, 166 IOps
  • read, 146.139MB/s, 3.089ms, 18705 IOps
  • write, 7.118MB/s, 63.971ms, 911 IOps

Full results here ; SSD Acceleration/Toshiba MQ01ABD100 with bcache