3Par has introduced its own specially designed storage servers. The company asserts that its design is better than an equivalent capacity SAN in a number of ways. 3Par CEO David Scott described the hardware and software to Techworld and explained several 3Par advantages.

The 3Par InServ X-Series Storage Server is like a SAN minus Fibre Channel switches. Connected host servers link to the InServ by Fibre Channel and connect to controllers. These are dual Xeon-based and link together via a full mesh backplane. They also use Fibre Channel (FC) to connect to storage arrays. The whole array can total 376TB with 2,560 147GB disk drives.

This hardware is driven by 3Par software on the dual Xeon CPU controllers which features clustering and virtualisation of the disks in its capabilities.

Controllers
Control nodes run 3Par's InForm operating software and it is this that drives the whole InServ system. A single InServ system can be a cluster of two to eight nodes with each node scaling from 4 to 24 2Gbit/s FC ports. Because of this modular style, InServ storage servers have a relatively low entry cost. With the scalability prices can rise; an SPC-1 benchmark-winning configuration with 8 controller nodes and 35.4TB of storage cost $1,413,798.

Controller node failure is handled by the failed controller's work being taken on by the other controllers in a failover routine using the backplane. Controllers have caches and these are battery-backed so contents can be saved to a controller's own disk if a power failure happens. A feature of the software is that control processing is separated from data movement. Data blocks are moved by ASICs with the Xeon CPUs handling the control processes. Volumes of data can be exported through any of the controllers.

Controller nodes are upgradable, David Scott said; "The P3 CPUs can be upgraded to P4 Xeons at 2.4GHz; the infrastructure can support higher performance. Take out the P3 node;plug in the P4 node; reboot and run much faster. The backplane is ready for the next (CPU) generation in a year or so. We would upgrade to 64-bit over time."

An S400 model has up to four controller nodes; a S800 up to eight.

Backplane
The backplane interconnects each controller node with dedicated 1GB/s links. 3Par decided to use a mesh backplane rather than either a bus- or switch-based design. Scott points out that, "The bus' problem is that it bottlenecks too easily. Switches alleviate the bottlenecks but are very expensive. They need to be protected against failure so you have to double them up in a SAN. We wanted to drive cost down and scale product down to the mid-range. If you try to scale switch architectures down then you run into cost barriers. 3Par is less expensive than deploying a SAN."

With the backplane interconnect the controllers form a cluster, a single cache-coherent system.

An InServ system comes in one or more 19 inch racks and these contain drive chassis', with the first rack also containing the controller nodes.

Drive Chassis
There can be up to 64 individual drive chassis; 4U rack enclosures with up to eight per standard 19 inch rack and nine racks in a maximally-configured system. Scott says, "A drive chassis has ten magazines with each containing four drives." The first rack in an S800 system would contain three drive chassis as well as the eight controller nodes.

Scott explains how the space on the drives is handled; "There is virtualisation across all the drives, fine-grained virtualisation. At level 1 a drive is split into fixed size chunklets of 256MB. At level 2 chunklets are aggregated into logical disks with (potentially) different properties: RAID 10 or 50; only 15,000rpm drives; whatever. At level 3 the logical disks are aggregated into volumes. The volume size is specifically requested by the aggregation system or a host. It obviates the need for volume management to take place in the hosts."

Scott asserts that this method is better for users, saying, "The arrays are really big error-handling machines. Read/write is the easy part. Does network-based virtualisation provide the same level of error-handling and reliability? Because, unlike 3Par, the total system is not under the total control of one vendor."

There is a cost advantage to 3Par as well, according to Scott; "There is massive under-utilisation of storage in arrays, around 25 percent. Apps and database admin people over-specify the storage capacity they need. In the (InServ) array we let the DBA and app think they have a terabyte but we only provision a small amount. We utilise at 90 percent, three times better, which drives capital cost down threefold."

Space and speed
He says 3Par is much better at copy-on-write snapshot technology too; "We use thin provisioning to solve a problem in C-on-W snapshot technology. Typically you would reserve space for the changed data in the copy to be written. But you may not need it. We use our thin provisioning and don't allocate space until the application actually writes new data. Our virtual copy utilisation is very efficient."

"We can make scores or hundreds of read-only or read-write snapshots of a volume. You can snap as regularly as you like. Many other arrays limit you to eight or 32 copies." Network Appliance and HP's EVA were mentioned in this context.

There is a recovery advantage as well; "You can do rapid application recovery by going back to a snapshot, mounting its read-write copy onto the application, and it's up and running instantly."

The InServ server is fast as well as good value; "Our system can sustain 100,000 IOPS, that's three to four more than IBM's ESS. That had a price/performance of $35/SPC-1-IOPS. We have a price/performance of $14.81/SPC-1-IOPS." These figures can be checked out at the Storage Performance Council.

The 3Par system is readily upgradable on a number of axes: more controller power; more storage capacity; more component redundancy. For businesses needing access to terabytes of data and valuing high data access speed and reliability the Inserv storage server is worth inspection.