The picture of customers buying storage that Adaptec has is that they prefer storage outside the server box. Storage is becoming an externally-attached resource and that brings different needs to the product. Should the disks inside the boxes be SCSI or Fibre Channel? Should they be the newer Serial ATA (SATA) or serial-attached SCSI (SAS)? The former provides SCSI-like performance with near ATA cost. The latter provides a way forward for SCSI from UltraSCSI 320.
Should the storage boxes be linked together as a single, and probably shared, resource? If so then the boxes need fast internal connectivity, a virtualisation software layer and, probably, information lifecycle management (ILM) facilities.
Mark Delsman is Adaptecs chief technology officer and he says that customers are not going to buy lots more SCSI and RAID arrays to go inside the server chassis; Data is moving from inside the server and going external. Servers and storage need to be networked together. The companys product range is changing to reflect this.
The external storage is being linked to servers either by Fibre Channel or iSCSI over Ethernet. There have been several iSCSI-based announcements. It doesnt mean Adaptec is biased; Were really neutral between fibre and iSCSI. There are separate portions of the SAN market: Fibre Channel in the data centre; iSCSI for regional offices and workgroups.
Adaptec is also pretty neutral about disk interfaces. Fibre Channel product is offered as are SATA products and SAS is also supported. Adaptec has an SAS-enabled enclosure into which you can plug SATA drives or SAS drives. This implies that there may be two tiers of storage in the cabinet. For example, there could be performance-focused drives and there could be slower but higher capacity drives providing some form of secondary storage.
Adaptec does have ILM ambitions and Delsman says, Well be getting into that in the future. How? We have three choices: partnerships; develop it; or buy it. We dont have a public strategy in that area yet.
The company believes that performance within an external drive array will need to increase. Asked about switching, Delsman says, Its something you will see in our future products, a low-cost internal switching fan-out. Sample product may start appearing in the second half of this year with a product launch in early 2005.
But switching doesnt mean SAN switches and directors. Adaptec is not in the Fibre Channel switch market; We do not offer switches. We view ourselves as at the host end and the storage end. We partner with switch vendors. We also offer partners host bus adapters.
Recently Adaptec bought Elipsan and its virtualisation technology. Delsman says, It was a spinout from Eurologic with software development targeted at iSCSI. It became a very embedded part of our strategy and its really enabling SAN features in our target storage. Were adding in snapshot and replication facilities to our software stack.
The idea is to offer a pool of storage across several enclosures using the Elipsan technology.
Intel dipping its TOEs in the storage water
Host server processor availability can suffer with iSCSI storage. Thats because the associated TCP/IP processing, needed to tunnel the SCSI commands and data across the IP Ethernet network, is pretty intensive.
Currently iSCSI has server interface cards which may or may not have TCP/IP offload engines (TOEs) on them to perform the TCP/IP processing needed. If the main server CPU does the necessary TCP/IP processing then the processor cycles available for other work decrease and users could well see a decrease in general performance, hence the offload idea.
Intel, saying that current TCP/IP offload is inefficient, is proposing the idea of having multi-cored processors with cores dedicated to particular tasks such as TCP/IP. In this case TCP/IP processing isnt offloaded at all.
Delsman thinks that Intel is highlighting how it, would like to play in this market. There are three ways of solving the TCP/IP offload problem. One is the Intel multi-core idea dont offload at all. Second is a partial offload. It needs quite a lot of hand-shaking. The TCP management part of the driver still runs on the host. I think this is what Intel is referring to (when it describes TOEs as inefficient). Third is a full TOE. Thats what we do. We interrupt back to the main CPU when the transaction is over. Thats the path weve chosen to go down Clearly there is much less hand-shaking going on between the TOE and the server CPU thus releasing the server CPU for its main tasks.
Lowering SAN costs
Dell has said that it is looking to offer lower cost SANs and that bought-in HBAs figure in its thinking. Adaptc is one of the two main HBA vendors. What does it think? Delsman points out that Fibre Channel SANs are limited in this area; With Fibre Channel SANs port costs are high. Qlogic talked about bringing down these costs with restricted features for Fibre Channel. It was a response to iSCSI.
With iSCSI there is a range of costs and options available to you. You could run a software driver through a standard Ethernet NIC and move up to a hardware-accelerated host adapter. Its not possible with Fibre Channel where everything is in the ASIC.
For Adaptec the road to lower SAN costs would appear to be via iSCSI and its use of what others call Ethernet economics.
What we are seeing emerge here is a line of thought that says iSCSI SANS are viable. That customers need SAN storage at a lower cost point than Fibre Channel, with inherent limitations, can provide. The enabling technology is iSCSI with full TOE interface cards, virtualised pools of mixed tier storage inside switched cabinets, and data being positioned on appropriate storage tiers within these cabinets. There will also be data security features such as snapshot and replication.
We might consider this as the commoditisation of SANs. Its not a term that Delsman used but it seems appropriate. SAN technology is being adapted and Adaptec sees itself having a large role to play.