alexnik108
Contributor
Contributor

Thin disk faster then thick eager zeroed

Hi,

any one can explain me why thin disk works faster then thick eager zeroed on same datastore ?

THIN:

fio --randrepeat=1 --ioengine=libaio --direct=1 --gtod_reduce=1 --name=test --filename=test --bs=4k --iodepth=64 --size=4G --readwrite=randrw --rwmixread=75

Jobs: 1 (f=1): [m(1)] [100.0% done] [284.3MB/96892KB/0KB /s] [72.8K/24.3K/0 iops] [eta 00m:00s]

test: (groupid=0, jobs=1): err= 0: pid=4133: Sun Jul 16 22:27:32 2017

  read : io=3071.7MB, bw=289577KB/s, iops=72394, runt= 10862msec

  write: io=1024.4MB, bw=96567KB/s, iops=24141, runt= 10862msec

THICK:

  read : io=3071.7MB, bw=172067KB/s, iops=43016, runt= 18280msec

  write: io=1024.4MB, bw=57381KB/s, iops=14345, runt= 18280msec

What a mystic?

Datastore located on 4x SSD, RAID: 0, VMFS6, ESXI 6.5

5 Replies
unsichtbare
Expert
Expert

Eager zeroed disks are tremendously more overhead for backend storage.

Is your storage deduplicating?

+The Invisible Admin+ If you find me useful, follow my blog: http://johnborhek.com/
0 Kudos
alexnik108
Contributor
Contributor

No, PERC 730 doesn't support Dedup.

0 Kudos
continuum
Immortal
Immortal

A thin provisioned vmdk reads from physical device and /dev/zero.
A  thick provisioned vmdk on the other hand always reads from physical device.
Of course reading from /dev/zero is faster than reading from a real device.


________________________________________________
Do you need support with a VMFS recovery problem ? - send a message via skype "sanbarrow"
I do not support Workstation 16 at this time ...

alexnik108
Contributor
Contributor

Bravo,

and everyone around is stupid ?

for that reason in all low latency whitepapers made by vmware they suggest for better performance to have thick eager zeroed disks ?

0 Kudos
continuum
Immortal
Immortal

it is not so easy 🙂
Thin provisioned vmdks will perform good as long as they are almost empty. The more data there is inside, the higher the fragmentation-rate will be so the performance goes down again.
To get a good AND predictable performance the best choice IMHO is to use eager zeroed vmdks on an unfragmented datastore.
But IMHO the main reason to use eager zeroed thick vmdks is the much higher reliabilty of thick provisioned vmdks.


________________________________________________
Do you need support with a VMFS recovery problem ? - send a message via skype "sanbarrow"
I do not support Workstation 16 at this time ...