[pve-devel] firsts cache benchs results

[Alexandre DERUMIER]

Here firsts results, 

It's seem indeed that writeback is the default cache mode (no cache option defined in qemu command line). 

bench: random write 4K buffered (no directio)
----------------------------------------------- 
#mkfs.xfs /dev/vdb 
#mount /dev/vdb /mnt/   (default xfs options)
#fio --filename=/mnt/test1 --rw=randwrite --bs=4k --iodepth=40 --size=1000M --groupe_reporting --name=file1 --ioengine=libaio 


virtio (guest kernel 3.6-rc4), disk cache mode in guest 
------------------------------------------------------------------------------------- 
no cache defined : cat /sys/block/vda/cache_type : write_back 
cache=none : cat /sys/block/vda/cache_type : : write_back 
cache=writeback : cat /sys/block/vda/cache_type : write_back 
cache=writethrough : cat /sys/block/vda/cache_type : write_through 
cache=directsync : cat /sys/block/vda/cache_type : write_through 


results: 

local raid controller (512MB CACHE with battery): 
------------------------------------------------- 
default: 15000 
cache=none : 8370 
cache=writeback : 15000 



nexenta - libiscsi: 
-------------------- 
default: 4000 (some hangs) 
cache=none : 14000 
cache=writeback : 5000 (some hangs) 

nexenta - scsi-host 
------------------ 
default: 6000 (some hangs) 
cache=none : 11000 
cache=writeback : 6000 (some hangs) 



rbd 
------------------ 
default: 6000 
cache=none : 6000 
cache=writeback : 7000 

sheepdog 
------------------ 
default: 300 
cache=none : 300 
cache=writeback : 300 




So for local storage, writeback seem really faster.
for iscsi, writeback seem slower (I have some hang during 5-6 seconds, maybe this is host flushing...?..)
for rbd, I don't have big difference. (I'll try to tune the "rbd cache max dirty age" parameter later
for sheepdog, difficult to compare, It's a small cluster..

I'll do more test with nfs tomorrow


Regards,

Alexandre