HyperFlex: An Enhanced Look

This is a post originally written for the company blog — posted here for posterity.

In the IT industry, the phrase “we are pretty much a 100% physical shop” is one that you dread to hear — especially from a fast-growing company. Such was the case with a leader in the financial services industry recently when they asked Sentinel to install a Virtual Desktop Infrastructure (VDI) solution for a new call center rollout of around 250 desktops as well as fully re-deploy their physical desktop and server infrastructures. They were pretty set on a hyper-converged solution and were looking for something scalable and easy to manage. To be successful, in the eyes of the business, the solution had to:

  1. Be solid. With internal hesitation to virtualization from the business, there had to be reliability.
  2. Be fast to deploy. To meet the aggressive deadlines, there could be zero delay on delivery or deployment.
  3. Be lightning fast. To aid in business buy-in and adoption, the solution had to deliver a better end-user experience than the current desktops. Performance was critical to that.

After reviewing the vendor options, the customer ultimately chose Cisco HyperFlex and VMware Horizon for their hyper-converged VDI solution. Aggressive deployment timelines were set and equipment was on the way. From there we moved onto the fun stuff.

The HyperFlex cluster was delivered quickly. Really quickly. Once the gear was on-site it was time to deploy. Before we go there, I want to touch on one particular aspect of the solution. Sentinel knows that maintaining data integrity and availability is essential to our customers as they adopt and adapt to new technology. How the Cisco HyperFlex solution delivers that can be summed up pretty easily:

  • The Cisco HyperFlex product line is a variant of the Unified Computing System (UCS) product line, and with that you have the full redundant design of dual fabric interconnects, full multi-pathing, and server hardware that is designed with zero single point of failure. In this particular deployment, we had four nodes (N+1) with dual fabric interconnects, and two 10GB paths from each of the HX240c nodes. Everything also ran on fully redundant power. It was a strong platform to begin from.
  • The SpringPath HALO Architecture is a file system — I am simplifying things here a bit — that allows for distribution of writes onto multiple solid-state drives (SSDs) across multiple nodes BEFORE acknowledging the writes. This maintains the data integrity by ensuring that there are multiple copies of the data on separate nodes in the cluster to prevent potential data loss.
  • The HALO Architecture enhances the data integrity by using a Log Structured Distributed Object Store to allocate the data as small objects across multiple servers in a sequential pattern, which are in turn replicated to other pool members to achieve data redundancy. By doing so, they increase not only performance, but the life of the flash layer disk in the servers as well as redundancy overall.

Back to the deployment. In a post on my personal blog, I mentioned that the HyperFlex deployment was pretty fast. Once you rack and cable the cluster, the HX installer is a breeze. What I love about the HX installer is the fact that it really does build the entire UCS deployment and makes adding a node to an existing cluster just as easy. Click. Click. Done. Overall, the deployment of the HX system after rack and cable took less time than installing the vCenter server that was required for the deployment (Note: The vCenter must be on separate hardware but can be moved into the HyperFlex cluster for ongoing operations).

After meeting the first two objectives, we needed to look at the speed. Since this was a VDI cluster, we made one small change (one line in a configuration file) to optimize the cluster’s L3 Cache for a read-heavy environment. Once that small change was made, it was time to run some tests. Since Sentinel doesn’t own the environment I will only include the following observations:

  • During testing of the 4-Node cluster with 4xVMs pushing I/O, the cluster achieved well over 125,000 I/Ops. Even in the worst-case boot storm of 250 users logging in within a one-minute period you would only really require 117,500 I/Ops, leaving plenty of room to spare. Keep in mind, this was not done in a controlled lab under ideal circumstances.
  • I was able to clone a 100GB (65 Used Thin) VM from template in less than three seconds. Seriously.
  • I deployed 250 linked clone desktops including two boots, customization, and domain join in under seven minutes. The bottleneck was the VDI limit on the maximum concurrent operations sent to vCenter (which I tweaked to 25) and probably the Active Directory domain join tasks as part of the customization. It was fun watching the vCenter task pane roll by so fast I couldn’t keep up with it.

The customer was extremely happy with the performance, scalability and easy management of their new infrastructure. The Cisco HyperFlex and VMware Horizon solution met the requirements so well that I better understand the hype around Cisco HyperFlex and the SpringPath HALO Architecture.

Of further interest in terms of scalability comes confirmation from Cisco that node capacity expansion beyond the current self-imposed limitation is in the works and will not be limited to hardware. External storage is also fully supported. This means you will have the capability to hyper-converge your core systems and still make use of external storage area networks (SAN) where business needs dictate.

All in all, HyperFlex is a rock solid platform with a fantastic and robust architecture that you would be wise to evaluate. Couple it with VMware Horizon for desktop deployment, and you have an infrastructure built to help your business achieve unprecedented levels of success. If you would like to learn more about HyperFlex or other converged/hyper-converged infrastructure solutions, please contact Sentinel for more information.