[zfs-discuss] Performance question

henrikj at henkis.net henrikj at henkis.net
Sun May 29 17:59:02 EDT 2016


Hi,

I’m comparing the ZoL om Ubuntu 16.04 with OmniOS to see it it is possible to switch for some of our systems. Running filebench against a simulated workload that we have does however show much higher latency for random reads with ZoL. Is this expected or could something else be at play here?

OmniOS r151018:
 2116: 1128.302: Run took 900 seconds...
 2116: 1128.302: Per-Operation Breakdown
closefile1           52850ops       59ops/s   0.0mb/s      0.0ms/op        8us/op-cpu [0ms - 0ms]
readfile1            52850ops       59ops/s   2.9mb/s     17.0ms/op      125us/op-cpu [0ms - 680ms]
openfile1            52851ops       59ops/s   0.0mb/s      0.0ms/op       26us/op-cpu [0ms - 0ms]
 2116: 1128.302: IO Summary: 158551 ops, 176.155 ops/s, (59/0 r/w),   2.9mb/s,   1415us cpu/op,  17.0m
s latency

Ubuntu 16.04:
 9550: 1451.634: Run took 900 seconds...
 9550: 1451.634: Per-Operation Breakdown
closefile1           23449ops       26ops/s   0.0mb/s      0.0ms/op       33us/op-cpu [0ms - 0ms]
readfile1            23449ops       26ops/s   1.3mb/s     38.2ms/op      300us/op-cpu [0ms - 961ms]
openfile1            23450ops       26ops/s   0.0mb/s      0.0ms/op       23us/op-cpu [0ms - 0ms]
 9550: 1451.634: IO Summary: 70348 ops, 78.155 ops/s, (26/0 r/w),   1.3mb/s,    812us cpu/op,  38.3ms
latency

Identical hardware, both pool consist of 8 mirrors, ARC limited to 1GB to simulate a larger working set.

filebench workload:
set $dir=/bench
set $nfiles=50
set $meandirwidth=20
set $meanfilesize=2g
set $nthreads=1
set $iosize=50k

define fileset name=bigfileset,path=$dir,size=$meanfilesize,filesizegamma=0,entries=$nfiles,dirwidth=$
meandirwidth,prealloc=100

define process name=filereader,instances=1
{
  thread name=filereaderthread,memsize=10m,instances=$nthreads
  {
    flowop openfile name=openfile1,filesetname=bigfileset,fd=1
    flowop read name=readfile1,fd=1,iosize=$iosize,random
    flowop closefile name=closefile1,fd=1
  }
}

Regards
Henrik


More information about the zfs-discuss mailing list