The founder and CEO of the second most successful iSCSI array startup, LeftHand Networks1, is trying to do it again. He's chairing the board of another iSCSI startup: Starboard Storage, which is joining the hybrid SSD-HDD array party today. What's wacky about this is that ex-colleagues of his at LeftHand have also started up NexGen Storage, which is trying to pull off the same same trick. Starboard is aiming at the SME market, claiming it has optimised both for performance and cost per GB.
The AC72 array, AC standing for Application-Crafted, is a rack-mounted, dual controller array, using Xeon Core5600 processors, in a 3U, 16-bay chassis. There are a few twists: a read and write caching tier of solid state storage is one – no RAID, and management from an application focus are others. Also it doesn't just support iSCSI, it offers Fibre Channel, NFS and CIFS access as wel: this is a unified storage box.
More ReadingStarboard Storage goes titsup: LeftHanders to leave the buildingUpstart Starboard Storage cuts staff, looks to flog bizHP iSCSI slogger usurped by younger flashy modelNexGen boss: If solid hardness isn't stuck into PCIe - forget itNimble not struck by Lightning: We're fine as we are, thanks
There are three STEC or Toshiba SLC SSDs included as standard plus twelve 6Gbit/s hard drives, either nearline SAS (24TB of 2TB, 7,200rpm drives) or 15K rpm performance SAS (7.2TB). The SSDS are organised into a mirrored pair for a 100GB write cache and a 200GB spillover read cache.
The base offering has a single controller with 24GB of RAM, and the option of a second one and 48GB of memory. There are two PCIe X4 and one PCIe X8 interfaces, dual 10GbitE (internal), dual 1GbitE ports for management, and IPMI with KVM over LAN support. Customers can buy 1 and 10GbitE and 8Gbit/s Fibre Channel expansion cards.
High-availability features include the dual, redundant controllers, hot-swap power supplies, fans etc.
The drives are aggregated by the Starboard Storage O/S into a single dynamic storage pool with the flash providing a read and write acceleration cache. It is part of a MAST (Mixed-workload, Application-crafted Storage Tiering) architecture and DRAM caching is also part of it. Hot, high access rate files are automatically placed in the flash accelerator tier.
Maximum capacity expansion is up to 474TB using 16- and 45-bay expansion shelves. As extra drives are added they are put into the pool and storage work is automatically load-balanced across them. Starboard says performance increases as drives are added, the extra spindles increasing overall I/O speed.
The software licence is all-inclusive and includes thin-provisioning, mirroring, asynchronous replication and writable snapshots. There is no deduplication.
Starboard says its AC72 uses storage applications. Lee Johns, the company's product management VP, said: "Storage Apps are the way we automate provisioning for customers to configure their shared storage. Today those Apps enable simple configuration of the Storage Pool, and volumes and shares for FC, iSCSI, NFS and CIFS. The Storage Apps will expand over time to include other storage infrastructure workflow like replication and migration as well as making it simple to create Application-Crafted Volumes and shares for key applications and virtualised workloads. The Apps provide a consistent methodology and therefore enable generalists to extend their expertise across NAS and SAN."
We asked Lee some questions about the Starboard product:
El Reg: How is the dynamic storage pool apportioned to applications?
Lee Johns: The Dynamic Storage Pool provides a single pool of storage with variable stripe sizes (Layouts) and data chunks on varying classes of disk (600GB 15K SAS, 2TB 7200RPM nearline SAS, etc) and each volume created has the SSD Accelerator Tier to call on as a caching tier. We have chosen this methodology because it provides immediate acceleration for hot applications with the minimum of management needed on the customers part. No pre-planning, no policy, no waiting for data movement to occur in a background window on the storage – just acceleration for the applications that need it when they need it. Because of the hooks we have in the Dynamic Storage Pool and the SSD Accelerator Tier we have the ability to optimise volumes for specific workloads.
El Reg: Is there application-focused quality of service?
Lee Johns: No. we do use I/O stream detection to determine random vs sequential loads and route them appropriately and we do prioritise in the system which data is getting access to the SSD Accelerator Tier at any given time but we do not require users to set QOS policy on applications. In general the users we are dealing with do not have the sophistication required for this. In my experience this is true of most customers. They simply want predictable high-performance for their apps when they need it. We have optimised for performance and simplicity.
El Reg: How is data protected?
Lee Johns: We use redirect on write snapshots and we have asynchronous replication. We do not have a traditional HW RAID controller. We create striped disk layouts that essentially provide the equivalent of traditional RAID protection but are statistically better in terms of availability since we can use free space to rebuild the layouts, and when Drives are rebuilt they rebuild much quicker. This also helps with overall utilisation of the drives.
Starboard and the competition
El Reg: How does this multi-tier, hybrid SSD and HDD array compare to roughly similar arrays from Nimble Storage, Nexgen Storage, and Tintri?
Lee Johns: The vendors you mention have taken a narrow approach to the customer problem. They have in the most part focused on virtualisation and iSCSI and pure performance. They are perpetuating silos. We have taken an approach of helping SME customers simplify and consolidate all of their workloads. We have FC, iSCSI, CIFS and NFS and we help them with unstructured, structure and virtualised workloads. … Fibre Channel is still about 60 per cent of SAN interconnects and 80 per cent of data is unstructured. Virtualisation is clearly becoming the dominant server workload and we handle that but, in terms of data, unstructured data is the biggest challenge for SMEs.
We are soon going to get new I/O like NVMe and SCSI Express driving even higher performance for servers and now the storage that is based on server architectures. We are entering a phase where performance takes a quantum leap and the challenge for customers is going to be how do they manage and consolidate storage for all of their workloads leveraging this next generation of performance. EMC is forcing customers through a transition with a mashup of two architectures, NetApp WAFL is 20 years old and HP, IBM, Dell and Cisco really do not have a mixed workload platform. Neither do the other startups. That is what we are focused on.
El Reg: What is the maximum storage capacity supported in a single AC72 array?
Lee Johns: 474TB. This is an important distinction with the Starboard platform. Because it is designed to handle unstructured data as well as structured and virtualised workloads it has high capacity. It is optimised for both $/IOP and $/GB. We provide 2 different disk Expansion Shelves. A 16 bay 3U redundant hot pluggable SAS Expander solution optimised for availability and a 45 Drive bay in 4U solution optimised for Capacity and density. Max config is achieved with 5 x 45 bay Expansion shelves connected to a single AC72.
El Reg: Can AC72 arrays be linked together? If so how and how many?
Lee Johns: No. We are not a clustered solution today.
El Reg: Is the accelerator tier a cache or a tier, meaning is the data in it also stored on disk or not?
Lee Johns: It is a cache. Data is also on the disk drives. This has benefits in terms of amount of SSD drives required, simplicity for the customer with no policy settings required, and immediacy of performance acceleration. [There's] no need to wait for a weekly policy to move data to the SSD tier in order to minimise impact on performance as you find in first generation architectures like Compellent where they recommend moving data weekly. You get performance when and where you need it.
El Reg: How is data promoted to the accelerator tier, in what sizes and how often. Ditto data demotion?
Lee Johns: Data is promoted on a constant basis. The chunk size depends on the layouts for the data. We have a variable chunk size from 16K to 512K. The write cache ensures that sequential data is sent direct to disk and the random data in the write cache is written to disk in background; as this is done the cache is refreshed. For reads the SSD acts as a spillover for DRAM cache. Once DRAM is full the SSD cache acts as a spillover. As this cache fills the older data is deleted creating space for new hot data.
Ending separate SME block and file silos
The point of this box is that there is no need for separate block and file storage silos or a forced choice of one or the other. It can do all of an SME's shared storage tasks while being affordable, fast, simple to manage, operate and grow – that's Starboard's pitch. The pitch is being accompanied by those from other hybrid array startups like Nimble Storage, NexGen Storage and Tintri, against a background of Dell and HP revving up their acquired iSCSI product lines: EqualLogic and LeftHand (the P4000).
The AC72 is available now, and is priced at $59,995, the manufacturer's suggested retail price. ®
LeftHand Networks was bought by HP for $360m. Competitor EqualLogic was bought by Dell for $1.4bn, making it more successful in sell-out cost terms.