Impact of different devices on ZIL performance

[Rand]

Just wanted to share some results i found while looking for a good device for a ZIL; haven't seen much in actual numbers while searching before.

O/C this is not necessarily reproducible for everyone so treat it as ballpark number; its not a scientific test either;)

Base array is a 4x2SATA3 Mirror (5940 RPM), i added the following:

-Intel SSD 510 120GB ~50MB/s
-Dell Perc 5/i with BBU/512MB Cache,
and an old slow 2,5" HDD/2 x250 GB 3,5 in Raid0 - both ~40-50 MB/s
-IBM M5015 - way slower than Dell Perc for some reason
-HGST SSD400M ~150MB/s

This last one is slightly slower than an Intel DC3700 so that should provide another boost.

All tests where run with large file copy initially and sync=always; I cross-checked the last with ESXi, VMWare Creation and CrystalDiskMark, albeit only on 1GB connection but got ~100 MB so i assume its valid.

The SSD400M is o/c not the most likely candidate but its currently obtainable fairly cheap in the bay over here.

I can run some more tests if anyone is interested and time permits

Edit: Will try to use a more complete approach as suggested below; this will take some time though.
Just forgot to mention those results above where with cifs but sync=always to simulate ESXi behaviour.

 

[joeschmuck]

I've done a little testing with respect to ZIL and L2ARC and have that posted here which originally started as an Intel vs RealTec NIC but also ended up including ZIL and L2ARC. It would be nice that if you want to post throughput results that you do a similar layout, meaning you completely identify your system and configuration so people can put into perspective the data you submit.

 

[Rand]

yes i've seen that, didnt know if you wanted other ppl to add to that thats why i created a new thread, but will add it over there as well

 

[joeschmuck]

Oh no, I didn't mean add to it but reference it as to what I did. I included all sorts of facts about my system and all the details on the testing so anyone could replicate my results. I feel that if you are going to throw up some comparisons about how a ZIL works that you should be as detailed in the testing as possible.

 

[Rand]

lol yeah got that on second read;)

 

[joeschmuck]

I think it would be good to add another person into the testing of a ZIL and it's performance increases, if any. Then draw a conclusion as to the real world performance increase as well. This would actually help people out when they decide if they want to add a ZIL or not. Since you read my thread, well you know my opinion already. Good luck and I'm sure you will have fun playing with it, I know I did when I did my testing.