Round Robin IOPS Limit

Software-based VM-centric and flash-friendly VM storage + free version

Moderators: anton (staff), art (staff), Anatoly (staff), Max (staff)

Post Reply
mphilli7823
Posts: 16
Joined: Wed Oct 31, 2012 7:41 pm

Wed Apr 29, 2015 6:35 pm

Does anyone know if it's best practice with Starwinds to adjust the default IOPS limit from 1000 to 1 in ESXi?

http://kb.vmware.com/selfservice/micros ... Id=2069356
mphilli7823
Posts: 16
Joined: Wed Oct 31, 2012 7:41 pm

Thu Apr 30, 2015 10:21 pm

wow no input....
User avatar
lohelle
Posts: 144
Joined: Sun Aug 28, 2011 2:04 pm

Fri May 01, 2015 9:16 am

I have always used 10 (and not 1). No special reason other than thinking it could be a little easier for the hosts and SAN?
User avatar
anton (staff)
Site Admin
Posts: 4021
Joined: Fri Jun 18, 2004 12:03 am
Location: British Virgin Islands
Contact:

Fri May 01, 2015 10:56 am

Sorry all ninjas are busy :)

General rule of thumb is faster CPU and storage back end you have smaller threshold value for path switch you set. 1,000 is definitely on a higher side and 1 exhausts system (especially hyper-converged setup...) so give a try to something within 10-100 range on your hardware. Your mileage may vary (c) ...

Good luck! :)
mphilli7823 wrote:wow no input....
Regards,
Anton Kolomyeytsev

Chief Technology Officer & Chief Architect, StarWind Software

Image
mphilli7823
Posts: 16
Joined: Wed Oct 31, 2012 7:41 pm

Fri May 01, 2015 2:47 pm

Cool...I'll change the values and do some testing and report what I find here...thanks!
User avatar
anton (staff)
Site Admin
Posts: 4021
Joined: Fri Jun 18, 2004 12:03 am
Location: British Virgin Islands
Contact:

Fri May 01, 2015 2:51 pm

Yup! Some sort of a table with numbers would be nice :)
mphilli7823 wrote:Cool...I'll change the values and do some testing and report what I find here...thanks!
Regards,
Anton Kolomyeytsev

Chief Technology Officer & Chief Architect, StarWind Software

Image
craggy
Posts: 55
Joined: Tue Oct 30, 2012 3:33 pm

Sat May 02, 2015 10:40 am

We always do round robin by bytes instead of IOPs and set it to 8972 so we can get maximum utilisation of each jumbo packet before the path switches.
User avatar
anton (staff)
Site Admin
Posts: 4021
Joined: Fri Jun 18, 2004 12:03 am
Location: British Virgin Islands
Contact:

Sat May 02, 2015 4:57 pm

Did you run any CPU and bandwidth tests to come up with a value?
craggy wrote:We always do round robin by bytes instead of IOPs and set it to 8972 so we can get maximum utilisation of each jumbo packet before the path switches.
Regards,
Anton Kolomyeytsev

Chief Technology Officer & Chief Architect, StarWind Software

Image
mphilli7823
Posts: 16
Joined: Wed Oct 31, 2012 7:41 pm

Mon May 04, 2015 1:24 am

Below is a screen shot of the results of my testing.

Test Setup
Test VM = Server 2008 R2, 8 vcpus, 16GB RAM, 40GB Drive
Hypervisor = ESXi 5.1, each hypervisor host has 2x 10GB nics for storage
Storage = Starwinds 6.0.6399, 2x Nodes 24GB RAM, 2x E5620
Storage network is 10gbit
Each storage node has a DataOn DNS1660D with 40x 600GB SAS drives + 4 100GB SSD as cache(LSI cachecade) connected to a LSI 9286CV-8eCC controller


I tested using PassMark's PerformanceTest 8.0 using the all disk tests. Each test was run twice and the averages were used.
As shown by the graphs the sequential read test was the biggest benefactor of this setting. Going forward I will use the iops=10 setting.
iops_round_robin.PNG
iops_round_robin.PNG (21.36 KiB) Viewed 13324 times
User avatar
anton (staff)
Site Admin
Posts: 4021
Joined: Fri Jun 18, 2004 12:03 am
Location: British Virgin Islands
Contact:

Mon May 04, 2015 11:00 am

Wow! That's a very impressive report! Thanks for sharing :) Just it case... Do you happen to have matching CPU usage on client and server?
Regards,
Anton Kolomyeytsev

Chief Technology Officer & Chief Architect, StarWind Software

Image
Post Reply