The hybrid cloud storage market has developed quickly, with at least a half-dozen software and hardware solutions now available. But placing these products into a production environment for primary or nearline data can be a daunting proposition, requiring the matching of workload profiles to the capabilities of the products.
In this article we look at the hybrid cloud products on the market today and attempt to assess how capable they are of dealing with the primary and nearline workload requirements of a typical IT organisation.
There are many different workload profiles in the data centre today. But, as a general rule -- and for the purposes of this discussion -- we can divide workload into two profile types: file-based and block-based.
File-based workloads will typically be NFS or CIFS data from end users. This includes Microsoft Office products and other local applications on user file shares. This data typically consists of large reads and writes, with files ranging in size from a few kilobytes to megabytes. Files are rarely reopened, and the working set of active data can be quite small. One study found more than "66% are reopened only once and 95% fewer than five times" in large CIFS environments.
Meanwhile, I/O response time for user data is less critical than for block-based workloads, and it’s unlikely the end users will notice an additional few milliseconds' delay in accessing their data. However, when organisations use NFS and CIFS for production workloads, it may create a different (and more demanding) I/O workload profile. A good example of this is the use of NFS for server virtualisation.
Block-based workloads are more complex. Latency -- the time taken to complete an I/O request -- is a key factor in the performance of block-based storage. Storage vendors have worked hard to reduce and improve the consistency of latency within their storage products.
Clearly, latency is application dependent, but it's impractical to deliver files across the Internet without local caching hardware. The ability to deliver block-based storage will therefore be highly dependent on the capabilities of the local caching device.
Hybrid cloud market analysis
Whether they supply physical appliances or not, hybrid cloud storage vendors don’t quote specific IOPS capabilities of their equipment. Even with dedicated hardware in place, local performance capabilities are difficult to predict because the ratio of local and cloud-based data isn’t always known. The vendor analysis in this article attempts to quantify the capabilities of hybrid solutions. As we will see, solutions include the use of solid-state drives and a minimum of Gigabit Ethernet (GbE) connections.
StorSimple. Caching appliances from StorSimple provide iSCSI block-based storage LUNs. The hardware models scale from 2 terabytes (TB) to 20 TB of local disk space with 400 GB to 2 TB of enterprise multi-level cell (eMLC) solid-state drives (SSD). Effective capacity is quoted as up to 100 TB, which is achieved by a mixture of compression and data deduplication. Total capacity (the local and cloud capacity combined) scales to 500 TB. All StorSimple appliances have four GbE Ethernet NICs, which are used for local and cloud connectivity. StorSimple appliances are well-suited to handling iSCSI workloads, with a mix of SATA drives and SSDs. This addresses the requirement of consistent latency for local I/O, scaling capacity to the cloud, and delivering to primary and nearline requirements.
TwinStrata. Hybrid cloud storage appliances from TwinStrata scale from 1 TB to 24 TB of local disk cache, using SATA and SAS drives. The systems support up to 256 logical LUNs at a maximum size of 384 TB and a total system capacity of 100 petabytes (PB). Network connectivity is provided by two GbE NICs in the base models, with two GbE and four 10 GbE interfaces in the high-end solution. TwinStrata also offers a virtual appliance, which has no limit on local storage, subject to a 50 GB minimum. The TwinStrata offerings are capable of handling primary workloads with SAS drives. The SATA drives are better-suited to nearline applications. The virtual appliance capabilities will depend on the resources provided to the virtual machine. One note of caution -- the iSCSI logical LUN size can be extremely large. Creating such large LUNs with a small amount of local cache could cause some performance issues.
Nirvanix. Nirvanix offers two NAS-only solutions: CloudNAS and Hybrid Cloud Storage as a Service. CloudNAS is a software solution that runs on the customer’s equipment. Deployment requires a minimum of 200 GB of disk space, and at least 100 GB is used as a local cache. Performance is entirely dependent on the hardware on which it is implemented. Nirvanix’s hybrid cloud solution is a managed hardware deployment at the customer’s data centre. The standard deployment consists of two nodes, each with a maximum of 200 TB using 2 TB SATA drives. Network connectivity is provided by either GbE or 10 GbE. The Nirvanix solutions are obviously targeted at NAS workloads but will not be suitable for more demanding NAS applications such as virtualisation. Care needs to be taken with software-only solutions to ensure they are deployed on hardware with adequate resources.
Nasuni. The Nasuni storage controller has recently added iSCSI to the existing NFS and CIFS file protocols. Each controller is configured with up to 12 TB of local disk cache. Nasuni also offers a virtual appliance, which can be configured with up to 64 TB of local cache storage. Because the Nasuni system retains all data for iSCSI LUNs within the filer, performance is more predictable but the scale of iSCSI LUNs is limited to the capacity of the device. Replication to the cloud provider is done using periodic snapshots to ensure that no backlog of unreplicated data builds up. This means the storage controller is capable of delivering to block-based and NAS I/O workloads.
Panzura. Panzura’s Quicksilver Cloud Storage Controller provides NFS- and CIFS-based cloud storage. The controller is available as a virtual machine that scales to 3.5 TB locally, or as a dedicated hardware appliance that scales to 324 TB of local storage with unlimited cloud capacity. The Quicksilver can also use up to 108 local SSDs. Connectivity for the controller is provided by two GbE or 10 GbE Ethernet NICs. Panzura has made efforts to ensure its hardware can meet local I/O requirements for CIFS and NFS workloads, and its use of SSD makes it well suited to more I/O-intensive NFS traffic.
Ctera. Ctera Networks' Cloud Attached Storage appliances provide up to 24 TB of local cache storage using 3 TB SATA drives. Each appliance has two GbE NICs for local and cloud network connectivity. The Ctera devices support NAS and iSCSI protocols for file- and block-based storage. The use of large SATA drives needs to be considered carefully in terms of performance. While these drives will work well with NAS file traffic, iSCSI performance may be less than desirable, especially in mixed workloads.
Hybrid cloud storage vendors are careful to talk about performance without quoting actual figures on the capabilities of their technology. Clearly, many variables are involved that make these solutions different from locally deployed storage. With iSCSI block storage, the solutions use different methods, including SSD (StorSimple) or local retention of 100% of the data (Nasuni).
Solutions from Nirvanix and Panzura are designed with significant capacity scalability in mind. This means back-end connectivity for these devices to the cloud needs high throughput. Nirvanix partners with a number of colocation providers to deliver dedicated bandwidth for this purpose. Panzura has included the option of SSDs for local performance.
Virtual appliance solutions need particular care. Performance is highly dependent on the resources assigned to the virtual machine. On the plus side, the use of virtualisation means resources can be added dynamically to boost performance when required.
Alternatives to providing dedicated bandwidth include use of WAN optimisation. This technology can be particularly effective on file-based data but may not be as efficient when accelerating optimised data from caching controllers. This is because WAN optimisation technologies look for repeated patterns of data or known data types, but the data being written to the cloud will be a mix of new files and file or LUN fragments and so will not be as easily optimised. Performance can be improved by increasing the size of the local cache, but this is clearly at the expense of the benefits of cloud storage.
There is a large variety in the design of today’s hybrid cloud solutions. Also, performance figures are largely absent. This is an area vendors need to address with sample workloads and performance metrics. As more deployments of this technology are made, the volume of available data should increase, providing customers with more meaningful figures on which to base their purchases.
Chris Evans is an independent consultant with Langton Blue, a London-based consultancy.