Performance testing

[Info@ipsafe.de]

Hello,

i am new to the forum but working with starwind for some years.
currently i am setting up new hardware for my HA cluster

with the new hardware i did some testing to find optimum performance
( OS 2012R2 , Mellanox 10gbe Network )

i found that:
- HDtune is not good for testing ( confusing results )
- ATTO shows better results

after testing with raid , i went to a nvme disk to eliminate the influence of raid settings.
after that i went to RAMdisk based testing.
.

Starwind RAMdisk

RAMdisk_.jpg (174.6 KiB) Viewed 3307 times

.
testing with slower disks showed allmost no difference with Jumboframes
with ramdisk, the Jumboframes showed about 12% better performance, allmost close to full wire speed.
but also std network setting shows a good iscsi performance.
( dont forget to enable also the switch to jumboframe ! )

after that I installed a RAMdisk (with Softperfect RAM Disk 3.4.8.0 free version )
and did further testing

on_RAMdisk.jpg (231.03 KiB) Viewed 3306 times

the local RAMdisk in direct access ( without Starwind ) shows a hughe performance of about 3000 MB/s

the Starwind disk ( .img on this really fast RAMdisk ) is showing only 40% of the performance ( in loopback mode )

what can i do to reduce this gap ?
- no antivirus is running
- testing with NTFS block size showed allmost no difference

from another post:
"...Also, take into account that StarWind adds virtualization layer to existing ESXi infrastructure, so 10-15% performance difference is considered acceptable."
so the loopback should show at least 2300MB/s

[Arcyne]

Hi, if you have high speed flash such as nvme you can connect more than one isci connection. Simply select the iscsi Target in the Microsoft iscsi initiator discovered targets and press connect and set up another local connection. You can leave it to the default round robin setting or perhaps change to lowest queue if that works best for your environment. In theory you can add quite a few additional connections until you have the performance / CPU balance that you want.

R

[Info@ipsafe.de]

Great,
2 iscsi connections bring much better performance ( only -20% loss to the direkt hardware access )
on smaler packets the gap is a little bit more.
3 connections or lowest queue setting have small positive effekt, so round robin with 2 connections is fine.
CPU load is 10-15% higher

this information should be found in every setup guide !

[Arcyne]

One additional suggestion. If you are running an AV, make sure you exclude the StarWindService.exe and/or the path to the FLAT file images. That had a significant impact on the CPU utilization if I remember correctly. (I am still exploring how to tune the service so YMMV.)

[Boris (staff)]

this information should be found in every setup guide !

As you may be aware of, environments of our users differ. We cannot say for sure that for some of them 1 loopback session is enough, while others are fine with 3 loopback sessions per target. This is generally more of fine tuning than just ordinary setup.
Nevertheless, we will consider your suggestion on introducing some information into the guides. Thanks for your input.