Tintri VMstore review: Fast as flash, cheap as disk

Tintri's hybrid array delivers superfast storage and supersimple storage management for virtual machines

speed light curves fast
At a Glance
  • Tintri VMstore T800 Series

infoworld editors choice plum

The world of storage is in the throes of a significant upheaval, as the increasing affordability of flash and the ubiquity of virtualization spark new and innovative approaches to the back-end storage array. Among the new storage companies at the forefront of this revolution is Tintri, whose founders came out of companies such as VMware and Data Domain. Tintri's VMstore appliance combines flash, disk, inline deduplication, and other software magic to provide cost-effective, high-performance storage specifically for virtual machines.

I tested the Tintri VMstore in a VMware environment, but as of Tintri OS 3.0 and 3.1 (which arrived in August and November, respectively) the VMstore also supports Red Hat Enterprise Virtualization and Microsoft Windows Hyper-V. In addition, Version 3.1 brings new capabilities for disaster recovery through tight integration with VMware's Site Recovery Manager, encryption for data at rest, and support for PowerShell scripting.

ReplicateVM and CloneVM are two Tintri capabilities that implement features also found within the VMware domain. (ReplicateVM is not part of the base product and requires an additional license.) The difference is that these features take advantage of the Tintri architecture to perform the replication and cloning tasks with maximum efficiency. CloneVM has the ability to create clones from current or past snapshots, as well as on a remote site. Similarly, SnapVM adds a number of features to the snapshot process with the ability to scale up to 128 snapshots per VM and thousands per data store.

Tintri architecture

At the heart of the Tintri design is a focus on the virtual machine, rather than volumes or LUNs (logical unit numbers), as the object of storage management. Management tasks operate directly on virtual disks, while monitoring is done at the VM level. This makes the VMstore remarkably simple to install and manage. The other key piece of the Tintri architecture is the company's patented "flash first" design, which entails writing everything to flash and attempting to keep hot data there so that all reads come from the flash tier as well.

At the highest level is a protocol manager that tracks all I/O to the VMstore on a per-VM and per-vDisk basis. This information is then used to provide quality of service to individual VMs. This makes it possible to run mixed-performance workloads on the same data store while delivering the performance required to each one. The Tintri OS applies specific performance enhancements like prioritizing a VM's access to a VMware swap disk to avoid performance hits when virtual memory limits are exceeded.

Tintri uses lower-cost MLC flash to get the best price per gigabyte of flash storage possible. This requires a more robust write algorithm to overcome some of the inherent problems of flash (and even more pronounced in MLC than the pricier SLC), including write amplification caused by the difference between the size of typical data blocks written to disk and the size of erasure blocks on the device. Other flash-specific optimizations include efficient wear leveling and garbage collection to maintain a balance between available space and read, write, and erase cycles. All data written to flash uses an inline data compression and deduplication process for the most efficient use of what gets written to disk.

tintri vmstore

The Tintri VMstore supports VMware vSphere, Red Hat Enterprise Virtualization, and Microsoft Windows Hyper-V. 

VMstore T800 series

Tintri's newest hardware offerings all carry a model number in the T800 series and deliver three different levels of capacity. The big difference between the models is in the amount of raw storage available in both flash and spinning disk. All three models maintain an approximate ten-to-one ratio of hard disk capacity versus flash. This is typical among hybrid systems. For example, both Microsoft and VMware use that same ratio when recommending system configurations for their Storage Server and Virtual SAN products, respectively.

Each Tintri appliance consists of a two-node server with enterprise-class CPUs and memory. Both nodes have access to the underlying storage hardware and function in an active-standby configuration. Data is stored on disk using a log-structured file system, meaning the VMstore does not use hardware-based RAID. (Instead, RAID6 is provided by the Tintri OS.) Tintri is not a converged system in that you don't actually run any VMs directly on the Tintri appliance.

On the software side, the Tintri OS runs a highly optimized Linux kernel with a number of open source components. The key to optimizing VM performance lies in analyzing the I/O traffic to each data store and identifying any potential performance issues. The Tintri system offers deep instrumentation and even an autodiagnostic feature that sends data back to a central site where further analysis can be accomplished. With this data Tintri can spot potential issues -- high IOPS, high latency, looming oversubscription -- and recommend solutions to its customers before the damage is done.

Many functions including replication and data movement are handled internally by the Tintri OS. Tasks such as cloning a VM are performed without any appreciable network traffic. This can be accomplished either through the Tintri management interface or through VMware vCenter using the VAAI (VMware APIs for Array Integration) functionality. For additional data security, you can purchase a VMstore with self-encrypting disks. These disks use AES-256 bit encryption and do not impact performance or capacity. Tintri also offers a software add-on for encrypting data at rest that works in conjunction with ReplicateVM.

tintri dashboard

Figure 1. The VMstore can be monitored and managed through VMware vCenter. 

Managing the VMstore

Simplicity is the name of the game when it comes to managing a Tintri appliance. While a simple dashboard gives administrators an at-a-glance view of overall system health, the secret sauce for all management of a Tintri appliance comes in the form of REST APIs. Thus, the management platform is agnostic regarding virtualization platform, as any solution must go through the REST API to do what it needs to do. The VMstore also provides a rich set of functions available for automation using your favorite scripting tool. In the Linux world the scripting language of choice is Python, while for a Windows-based deployment you would use PowerShell.

Tintri resources can be managed from within VMware vCenter as well. Figure 1 shows the vSphere Web Client with the Tintri performance graphs and information displayed. From this view you can quickly grasp overall performance along with the impact of individual VMs on the system. A Tintri option under the Manage tab allows you to enter credentials for the VMstore and the vCenter Server, as well as configure and monitor the default snapshot schedules.

VMstore performance

For my testing, I was provided remote access to the Tintri Lightning Lab with three Dell PowerEdge R270 servers playing the role of vSphere hosts, each with 128GB of memory and two Intel E5-2620 CPUs. Each PowerEdge system had at least one 10GbE network connection to various Tintri VMstores. The lab included a VMstore T880, a VMstore T620, and two VMstore T540 systems (see Figure 2). As in my review of VMware's Virtual SAN, I used the VMware I/O Analyzer virtual appliance to simulate different workloads.

tintri lightning lab

Figure 2. My test environment included three VMware vSphere hosts with 10GbE links to VMstore T880, T620, and T540 appliances. 

I used the same Max IOPS workload to examine the impact of multiple VMs and multiple hosts on VMstore performance. A single host with four VMs averaged a little less than 30,000 IOPS total while the same host with eight VMs moved the number up to around 35,000 IOPS. Moving to two hosts with four VMs each bumped the number to a tad shy of 64,000 IOPS. Three hosts with four VMs each pushed the total to right at 75,000 IOPS. All of these tests were performed on the newest T880 host. Similar tests on the T620 resulted in somewhat lower numbers.

Pricing for the VMstore T820 starts at $74,000 and includes 1.5TB of flash storage and 20TB of raw disk space. The lower-end VMstore T820 comes with 1GB networking, while the two higher-end models ship with 10GB network cards. List price for the VMstore T850 with 5.3TB of flash and 52TB of raw disk space is $160,000. List price for the top-of-the-line VMstore T880 with 8.8TB of flash and 78TB of raw disk is $260,000.

At the lowest end, the VMstore T820 represents a significant value for organizations looking to beef up the performance of their VM storage. While the total raw storage for the VMstore T820 might not sound like much, the effective storage after compression and deduplication can be as much as three to five times the raw capacity.

Tintri's flash-first, VM-oriented approach to data center storage has produced a high-performance storage product that not only keeps costs in line with traditional disk storage arrays, but requires a minimum of effort to install and manage. Tintri's deep monitoring provides the means to track system performance and identify any potential problems posed by oversubscribing the available flash. The simple addition of nodes scales up both performance and total storage in a seamless fashion. The addition of support for Microsoft and Red Hat virtualization brings a new dimension of flexibility to this product line and broadens the potential customer base.

InfoWorld Scorecard
Availability (20%)
Interoperability (10%)
Management (20%)
Performance (20%)
Scalability (20%)
Value (10%)
Overall Score
Tintri VMstore T800 Series 9 9 9 9 10 9 9.2
At a Glance
  • Tintri's VMstore combines MLC flash and SATA disk drives to provide cost-effective, high-performance storage specifically for virtual machines.

    Pros

    • Supports VMware, Microsoft, and Red Hat hypervisors and mixed environments
    • Offloads storage tasks from hosts via VMware VAAI and Windows ODX
    • Provides QoS on a per VM basis
    • Integrates with VMware vCenter
    • REST API supports automation through Python and PowerShell scripts

    Cons

    • Performance drops significantly if you exceed SSD capacity
    • Dynamic monitoring requires connection back to Tintri central

Copyright © 2015 IDG Communications, Inc.