Tag Archives: storage

Cohesity and Veeam, better together.

At the end of last year, my Veeam repository storage was filling up and the storage was also a very old FC storage that I wanted to get rid of. My CTO tasked me to find a solution that would be cost effective and would scale as our backup storage needs were growing at a considerable rate. The storage would be placed in our secondary site where our Veeam copy jobs would reside.

For you to get an over view picture of our setup, here is a simplified diagram where our Kopavogur site is our primary site, and Akureyri site is our secondary site. We have more proxies, VMware hosts and some guest interaction proxies at customer locations that we back up to our system, but that’s out of scope of this blog so those components are not shown.


In our primary site I do the daily backups to an on-site repository for quick recovery of VM’s and data, but all copy jobs from the site goes to our secondary site in Akureyri.

When looking for solutions for the Akureyri site, I evaluated the benefits of using a deduplication appliance since copy jobs can greatly benefit of such storage as the same VM images are saved over and over again and in our backup policies that meant 10-15 times depending on the level of protection the customer wants. My conclusion that a deduplication appliance would be a great fit for the job, but my previous experience with deduplication appliance where performance would not scale well and fork-lift upgrades were need to add more performance of the solution did create a show stopper in that route.

Then I got news of Cohesity, as my friends Frank Brix Pedersen and Paul Schatteles left PernixData (RIP Pernix… L) and started working for a startup company Cohesity. I got an introduction of their product at a local VMUG event where I’m a leader, and also a live demonstration of the product at VMworld Barcelona in 2016. There I met other Cohesity staff as well so I got a good insight on what their plans were and how they were creating the solution out of hyper converged nodes where compute performance would scale linearly with the storage capacity. This got my attention as this method would get my wishes for a deduplication appliance back on track.

My next step was to get a POC underway to evaluate the solution and we agreed on pre-defined goals for the 30 day POC to be successful. I listed up 9 different points that would qualify as a successful initial setup, and another 7 points in regard of functionally and performance requirements.

Some of the initial points were (points were more detailed in the actual POC document)

  • Successful base installation of the product in Akureyri Site
  • Alerts and call home functionality
  • Initial setup of View box and views to publish CIFS share to the Veeam repository server
  • NFS share creation and connection to my ESXi hosts in Akureyri site and creation of NFS storage for archiving

Functionally and performance requirements points (again points were more detailed in the POC document)

  • Maintain data availability throughout a simulated hardware failure (node reboot/, drive failure etc…)
  • Successfully function as a Veeam repository target for the POC timeframe.
  • Successfully be able to perform Instant recovery with Veeam at a reasonable performance.
  • Stable user experience throughout the POC timeframe
  • Achieve a dedup ratio of 1:5 at the end of POC timeframe
  • Successful support case generation and acceptable response time from Cohesity support.

I got the box in mid of December 2016 and with help from Frank Brix I installed and setup the solution remotely in the Akureyri site. After some initial testing I connected Veeam to the box and everything was up and running the same day. As the holiday season was started I was unable to start the POC work right away, but at the end of 30 days in early January every point in the POC document was fulfilled except the 1:5 dedup rate. That was due to the fact that only 3 weekly copy job runs had run in the timeframe, but as my dedup rate was almost 1:3 I concluded that the theory worked. I would get a 1:5 dedup rate after 2-3 weeks.

We decided that the POC were a success and went ahead with the purchase. After few weeks, as projected I got my 1:5 dedup ratio that was defined in the POC document so my CTO got his Cost per GB projection confirmed and we went on with our daily business.

Now few months later, I’m really happy with the product. I have created a few minor issue based cases with Cohesity support, and always got great response time and help on my issues. Dedup ratio is on the rise, and in the 4 node, 2U box I have more than 400TB’s of logical data stored at this time, and I except to have this number doubled in the next 4-6 months without having to purchase another node. – Nodes can be added later when needed and the great thing about expansions is that I have a linear expansion path on both storage, performance and cost.

Last week I did a case study document with Paul that can be downloaded here from the Cohesity website where we wrote about the project.

Share or comment if you find this article helpful



Samsung 950 Pro M.2 512GB vs Mushkin Scorpion Deluxe 480GB

I got myself a new “disk-drive” today for my home workstation, if you can call a M.2 NVM chip a disk-drive!

I wanted to get a new drive for my OS and programs, as well as I use my workstation for GIS work, where I have both many, and large files opened in my GIS application. I also do graphics work on this PC using Lightroom, which uses huge number of index files for my photo collection.
What I got was a brand new Samsung V-Nand 512GB 950 Pro M.2 NVM Express, and I decided to test it against my older Mushkin Scorpion Deluxe 480GB PCIe based drive.
I wanted to find out how the new drive would hold up against my older Mushkin that I have been using for several years now. I used FIO to run several tests, 4K and 256k block size, Random and sequential reads and writes, to get some different views on the drives performance.

iops-4kUsing 4K block size during a 60 sec testing period, I get premium performance from the Samsung drive and pretty consistence performance though out the different test. The Mushkin delivers great performance as well during sequential reads and writes, but suffers when doing random writes and reads.

iops-256kUsing 256k block size, the Mushkin shines! The PCIe bus based drive delivers more than twice the performance of the Samsung M.2 Based drive.  The Samsung is like in the first test, more constant though.

My initial thoughts were to stop using the old Mushkin drive on my workstation, and move it over to my VMware server for PernixData FVP cache, – but now where I see the throughput difference for large IO’s on the Mushkin drive, I believe I’ll use that one for my GIS application files, and the new Samsung drive for OS, programs and Lightroom catalog files. I guess my VMware server has to stick with a normal SATA based SSD for now…



VSAN 6.0 in a nested ESXi 6.0 lab

I wanted to test VSAN in my lab without having to go out and buy SSD’s or invest in more hardware in my lab

The obvious path was to spin up several ESXi VM’s and do the settings in regard of networking and set normal HDD based volume as SSD. And to make things easy for you I took screenshots and wrote down every setting and step I made in this blog post.




To prepare networking for the ESXi VM’s you have to set Promiscuous mode to “Accept” in the security settings for the portgroup you place your ESXi VM’s on. You should not do this in a production installation on your whole vSwitch. In my lab I created a “NestedESXi” portgroup, where I enabled promiscuous mode by overriding the vSwitch default setting of “Reject” a VMware KB article explains this a bit more

This allows packets to travel from your physical nic on your ESXi host, up to the virtual nic of your virtual ESXi host, and up to its virtual VM’s virtual nic. Think inception + communications between each state.

Next thing to do is to create the ESXi VM’s. Select “Other” in “Guest OS Family”, and select “VMware ESXi 6.x” under “Guest OS Version.

This is pretty straight forward, but there is one setting in the “customize hardware” tab, and that is the option to set “Expose Hardware assisted virtualization to the guest OS” under CPU section.

Other settings on the VM’ is 2 cpu and 16GB of RAM (VSAN 6.0 memory requirements) state that each host should have a minimum of 32GB memory to accommodate the maximum number of 5 disk groups and a maximum of 7 capacity devices per disk group, – but in this lab test where I will only present 1 SSD and 1 HDD to the VSAN cluster, 16GB for the ESXi VM should work fine.

For the disks, I add one 4 GB disk for ESXi Installation, one 50GB disk to act as a simulated SSD disk, and one 150GB disk to act as a capacity device

I also in this step I select the “NestedESXi” network port group I prepared earlier.

I created 3 identical ESXi vm’s like this, and on more that had no extra hard disks, to test out the remote storage access of the VSAN cluster. VSAN requires 3 hosts as a minimum, with minimum 1 flash device and 1 spinning disk.

Next thing is to add the VM’s to vCenter as ESXi hosts.

I had earlier assigned IP addresses and created DNS records for the ESXi vm’s and I added the new hosts into a folder just for housekeeping reasons.

Before I create the VSAN cluster, I have to prepare the 50GB Hard disks and mark them as flash disk. In vSphere 6.0 this is really simple, just select the disk device and click the “F” button

This gives me a confirmation dialog to mark the selected disk as flash disk, and there you hit “yes”

This will mark the drive type to Flash

I also have to prepare a VMkernel port for VSAN SAN traffic. In this lab I’ll use the default vmk0 adapter for both management, vmotion and VSAN traffic. In production you should separate this though.

I do this for the other 2 ESXi VM’s and now everything is set up to create a VSAN Cluster.

To enable VSAN, select the “Turn on” checkbox under “Virtual SAN”

And then add your nested ESXi Hosts to the cluster.

After a minute or two, all the disks for the nested ESXi hosts automatically joined the VSAN and created a vsanDatastore.

And that’s it! – Now I have a VSAN datastore in my nested ESXi cluster.

As this is nested, using “fake” flash devices, I don’t expect to get much performance out of this, but for testing the process of creating a VSAN cluster this setup works great.

I hope you like this post, and send me your thoughs in the comments or on twitter.


Storage in the home lab.

Home Labs in general

When asking my colleagues what to run as a storage platform in my home-lab, I got an honest question from a fellow blogger and vExpert Rasmus Haslund (@haslund)

What are your requirements, challenges and constrains??
My answer: “Well, I want all the features and best performance, but I have limited or no budget!”

This could easily be applied to your production setup where you have the challenge of providing a stable service level, while having limited budget on external storage. So if you work for a small/medium company looking for a storage solution for your virtual workloads, read on and hopefully you can apply the solution described in this blogpost to your installation.

The challenge

As a vExpert, blogger and enthusiast for all sorts of storage and virtualization solutions, I find it necessary to have a lab at home to do tests and evaluate different solutions. I also run several vm’s for my home network that I have to take care of and have to answer to my son and wife if I screw up!

For quite some time I had a limited flexibility in regard of the lab and to maintain some level of service for my home network I had to find a better solution.

My son has a Minecraft server running that need to be up in the evenings specially, and my wife’s ideas about SLA for her e-mail and picture library in this regard is that a 100% uptime is “normal”!! So it’s tough ground to maintain and also have flexibility when it comes to testing and running some ad-hoc workloads.

In my basement there is a storage space and after I got a networking cable down there from my apartment on 2th floor, I could start up more hardware without my family being disturbed by noise and cables running all over my desk. Down there I can maintain a stable setup for my home network and have some extra hardware to play around with when I need to try out something.

When I got the chance to repurpose some servers from work I decided to redesign the home lab. It had been running from a one ESXi white-box host with 1 x Intel I7 3770K CPU and 32GB RAM and surly could befit of more CPU and RAM resources.

To set out some requirements and figure out the challenges.

The goals

  1. Maintain reasonable level of uptime and performance of my home network.
  2. Have available disk space and resources to set up a nested ESXi environment for testing different setups and solutions without exposing the home network to risk.
  3. Have a storage solution to be accessible by my 2 ESXi hosts.
  4. Minimizing heat generation and electricity cost for running the home network, but still have the ability to spin up more workloads for testing in the lab when needed.

The hardware

The servers I got for the lab are pretty massive!

3 x Dell PowerEdge R710, each having dual X5675 3,0Ghz CPU’s and 288GB of RAM. Each server has 4 x 1Gbit network cards onboard, and 1 x dual port 1Gbit NIC. Each server has the Dell H700 SAS controller (LSI based controller)

The solution

When looking for a storage solution I decided to use one of the R710 machines as an iSCSI target device as it had 6 x 3.5” drive bays. There I could place my 6 x 2TB SATA drives I previously had in my white-box server. This R710 server would become the shared storage for the 2 ESXi hosts as well as being a proxy server for my Veeam Backup installation, a Minecraft server for my son and a PLEX media server for my home entertainment system. (All those workloads that had been running on my wife’s desktop for some time, much to her enjoyment as you can believe) On one of the ESXi hosts I would run my home network workloads, but have the option to turn on ESXi host 2, and for lab testing.

I looked at several options, both Linux and windows based, virtual and non-virtual, that would enable me to run both the NAS iSCSI workload, but also the Veeam proxy, PLEX and Minecraft service. The setup I found most appealing for testing the different RAID levels and was a non-virtual windows based Starwind Virtual SAN solution

The main reason for running the workload in a non-virtualized Windows installation, was the fact that this enabled me test different IO and cache policies on the physical volume used as an iSCSI target. On native windows I could use the LSI MegaRaid Storage Manager to create and destroy volumes without having to reboot the server.

At a later stage I might run ESXi on this host, reducing the footprint down to 2 physical R710 machines using Starwind 2 node cluster setup.

Features of the Starwind SAN solution that I found interesting

Main Product page and Free Product Page

There are several features in the Starwind software that I found extremely cool. Also the simple setup and configuration process of the solution is truly remarkable. It makes testing the different configurations fast and easy.

To name a few features that got my attention while testing the software, that other users could benefit of both in regard of lab testing and for production workloads.

  • Use of defined amount of RAM for cache for each defined iSCSI device.

This allows me to define the amount of RAM assigned for the NAS storage role, keeping RAM available to other workloads on the server. This also allows me to define different devices and iSCSI target with different amount of RAM depending on workload types. Keep in mind that if you assign many GB’s of RAM for cache in a production setup, make sure you have a UPS to be able to commit all cached writes to disk!

  • Create a RAM based disk device.

Using this super-fast iSCSI target is great for testing and deploying temporary workloads in the lab. I plan to experiment with this feature more, but keep in mind this in in memory, so data is not written to any persistence storage! Non-persistence VDI disks (linked-clones) come in mind or classroom VM’S could use this feature to give great end-user experience.

  • Log-Structured File system while thin-provisioning the storage device.
    This feature turns otherwise “all writes are random” situation while running mixed virtual workloads, into sequential write on the underling storage. A whitepaper (https://www.starwindsoftware.com/whitepapers/eliminating-the-io-blender-by-jon-toigo.pdf) by Jon Toigo explains this in great detail, but this features boosts the benefits of thin-provisioning to a whole new level!
  • Publish a physical disk directly as an iSCSI target.
    This feature caught my eye, and I still have to investigate the pros and cons in this regard.

 The Network design

To give out a clear picture of my setup, I made the following diagrams.

Layer 1 Diagram

Picture 1: Cabling layout

  • 2 x 1Gbit network interfaces are connected from each ESXi host to the iSCSI NAS host.
  • 2 x 1Gbit network interfaces are used for vMotion and replication.

Layer 2-3 Diagram

Picture 2: Layer 2-3 diagram

The diagram shows the networking layout of the 2 iSCSI networks. Different subnets are used for each physical adapter assigned to iSCSI to provide active-active paths to the iSCSI target machine.
Path selection Policy is set to “Round Robin” for link load balancing

vMotion network between the hosts are bound to 2 physical network adapters, on a single subnet.

Storage design

For testing purposes, I decided to install Windows 2012 directly on a 2 disk mirror, and have the 4 extra drive slots to test different RAID levels and drive types. This allowed me to run the LSI MegaRaid Manager software and set different settings on the volumes and save me the reboot time when changing raid levels or drive types.

I had 4 x 2TB, 7.4K SATA drivers and 4 x 600GB, 15K SAS drives to test.

 Different Raid Levels and drive types.

First I tested out different RAID levels and on both types of drives, and ran FIO tests locally on the volume created.

Different Raid Levels

It caught my eye that when using the SATA drives, performance gain from Raid 10 to Raid 0 was minimal, while the SAS drives had huge performance gain while running Raid 0 vs Raid 10. Later I plan to do a 6 x 2TB SATA drive Raid 10, and that’s most likely the configuration I’ll end up using for my lab setup.

For the remaining of the performance tests, I ran the Raid 10 setup on the 4 x 15K drives, and the main goal was to find out if the different deployment options on the Starwind SAN software made any measurable difference, and also to see how it performed against the native Windows 2012R2 iSCSI target.

CrystalDiskMark tests

First test was done by using CrystalDiskMark measuring MB/Sec

CrystalDiskMark MB/secCrystalDiskMark IOPS

The tests show that in any configuration, the Starwind SAN software outperforms the Windows 2012R2 Built in iSCSI target solution by far. The only tests where the Windows iSCSI target was close was while testing sequential reads or writes, and I believe the limiting factor was the single threaded process and use of one network connection between the 2 physical machines.

All the random reads and writes tests showed huge benefits while using the Starwind solution. The CrystalDiskMark is a simple tool to test disk performance and it does not allow you to change from the fixed 4K block size, or go beyond the queue depth of 32.
The H700 controller on the iSCSI target machine has queue depth of 975 and to utilize the 2x 1GB network connection I moved from the CrystalDiskMark to more customable test tool, FIO.

To create a baseline and to get the maximum performance without the limitation of my 2 x 1GB network connections between hosts, I ran all tests both locally on the iSCSI target machine and on a remove VM. To test the performance running locally, I mapped a set of iSCSI targets as drives on the windows iSCSI target machine and an identical set of targets to my ESXi host.

The FIO test setup.

Each Starwind iSCSI target configured with 10GB Memory Cache

VM runs on a ESXi 6.0 Hosts, connected by 2 x 1GB Network cards, each configured on separate Subnets, – Round Robin PSP selected

FIO WindowsIO Engine settings:
Random Read/Write:    33/66
Block Size:                          64K
Queue Depth:                  975
4 x 15GB Jobs, 4 files each

FIO MB/sec


Direct = FIO Run directly on iSCSI target machine disk volume
Flat = Starwind iSCSI Target with Flat provisioned Image file
LSFS = Starwind iSCSI Target with Thin provisioned disk using LSFS
LSFS Dedup = Starwind iSCSI Target with Thin provisioned disk using LSFS and Deduplication enabled
Physical Disk = Starwind iSCSI Target from physical disk

The direct testing showed how much performance I could get from direct disk access. As I ran those tests, I got a clear picture of the different deployment options in the Starwind SAN software and my findings showed that the Thin Provisioned disk utilizing the LSFS was the fastest option.

While testing deduplication, performance dropped to some degree compared to the LSFS option in regard of IOPS. I also noticed some (5-7%) increased CPU load on the iSCSI target machine while I was running the tests. Also keep in mind that each 1 x TB of deduplicated storage requires 3.5GB of RAM. In my setup this was not an issue but if you have limited amount of RAM you should take note of this fact.

Future plans and few points.

Later, when I have finished the performance tests, I plan to create a target device, for the system drives for my home network VM’s, using deduplication, and save space there, but I’ll leave that option disabled for the PLEX media library and also the photo library as those media files are unlikely to be good candidates for deduplication.

When rebooting the iSCSI target machine, I noticed that the FLAT file and Physical DISK targets were active soon after boot time, but the thin provisioned LSFS and LSFS Dedup targets took some time to become active. After some investigation I saw the LSFS files were all read though, most likely due to file-checking and verification. My test targets were all 100GB in size and it took some time (5-10 minutes) to become active. When evaluating the benefits of FLAT or Physical targets, I guess if you have large targets (3TB as in my case for PLEX media library) you would prefer to use the FLAT file option there to have the targets online soon after reboot.


For a 2-3 hosts setup like mine, or even 1 host installation, it is clearly beneficial to use the Starwind SAN iSCSI software rather than direct disk access or native Windows iSCSI target software.

My findings on different deployment options will hopefully help you decide on what to go with both in your lab or production installations.

A colleague of mine pointed out that my home lab had more performance than many of his client’s production setups, and told me that if I was happy with the performance of the Starwind SAN software, he could recommend it to his clients for production!