IP Stack Location Might Rush Gigabit Ethernet To SANs.

Combats the 'unreliable network assumption'

In the continuing saga of transport interfaces for SANs, a serious contender to SCSI and Fibre Channel is Gigabit Ethernet. GE is an extension of the IEEE 802.3 Ethernet standard, developed about three years ago. E originally targeted 1Gbps, and is now looking to 10Gbps in short order. Strictly speaking, GE is a combination of Ethernet and Fibre Channel, which permits backward compatibility with Ethernet and incorporates some of the efficiencies that Fibre Channel offers.

It would sound like the optimum plan, but there are some hurdles to overcome. GE is a shared protocol, opening the network line to all devices that might need it. It, therefore, does not guarantee that the bandwidth is always available. Fibre Channel, on the other hand, makes such a guarantee. Further, the size of an Ethernet message is 1,518 bytes. FC allows special hardware to communicate at about 128MB in a single sequence.

Yet a telling difference between the two technologies is that GE retains what is called the "unreliable network" assumption against which Ethernet was designed. This merely means that additional software controls are required to handle network failures. In a controlled and reliable network, these are not needed. In order for the steps to be performed, but at a faster rate, the protocol stack needs to be relocated into hardware.

On The Hardware

Placing the IP protocol stack in hardware rather than software would dent the unreliable network assumption, but not all technologists are lining up. Internet experts are against moving TCP/IP from software on general-purpose microprocessors into special-purpose silicon. These experts suggest that TCP/IP software has no hot spots (frequently executed code) that would speed up significantly in silicon.

Simon Fok at NetConvergence finds this curious. He says: "I cannot understand why anybody would object to have TCP/IP implemented in hardware in one form or the other. There are already TCP/IP chips available and integrated to 'small' Web servers. To find them, simply do a search on TCP/IP + Web and you will see them. It's that common already. From the viewpoint of IP storage and, specifically, for iSCSI, the performance will never be close to wire speed without any hardware assist in TCP/IP."

One company stepping up to the challenge is Silicon Valley start-up, Alacritech, founded by Larry Boucher of Adaptec and Auspex fame. Alacritech is launching a new line of NICs (Network Interface Cards) based on an architecture called SLIC (Session Layer Interface Card),

SLIC takes over a large part of the IP stack using a custom ASIC known as the Internet Protocol Processor. The processor is part of a NIC with four ports of 100Mbps Ethernet running wire-speed TCP/IP with reduced server loads. Much of the performance improvement provided by the technology is driven by reductions in CPU interrupts. Conventional NICs bounce data between buffers and CPU memory. SLIC takes one transaction to load the data.

Troika Networks puts the IP stack into a multiprotocol board to run over Fibre Channel. SAN2 Controller 2000 is capable of communicating in the protocols FC-SCSI, FC-IP, and the very new Virtual Interface architecture. FC-SCSI and FC-IP are protocols needed to seamlessly share data and communication among storage and servers. FC-SCSI is a communication path optimized for backup and storage management. It also empowers the server OS to talk to the storage subsystem disks and vice versa. FC-IP is the protocol for clustered servers and shared server network communications.

Intel, Microsoft, and Compaq developed VI or Virtual Interface. VI assures that high-performance applications have fair access to the communications path. It gives high-speed, low-latency applications, including database software a separate communications path that bypasses the network stack.

NetConvergence's Fok has looked at both approaches. He notes: "From the point of view of a third party, both approaches are very proprietary and not easy to adopt. Fortunately, with the advances in Network Processors available from Intel, IBM, Lucent, and Motorola, to name a few vendors, one will be available to get TCP/IP implemented at hardware speed and, at the same time, allow for adaptation for other applications. This approach is flexible, as well as high adaptive, portable, and open standards-based."

CNT was one of the earliest companies to address the use of IP as a storage transport. CNT's Brian Larson commented to CTR: "At this time, CNT has the IP stack directly on the processor boards that support the associated 110 (i.e., 10/100 and 1 Gig Ethernet). In addition, CNT's Director platform is a distributed system using independent processors. The data paths are established through a high-speed midplane switch and IP and used as the communications vehicle within the architecture."

Larson continues: "The use of IP was selected for flexibility and standards. Why reinvent the wheel when this protocol serves the purpose? As TCP/IP is put down on silicon, a new approach may be used, but at this time, we don't see it mature enough to make any significant modifications to our existing approach."

GE SANs?

The location of the IP stack matters to storage because GE is a serious contender for interface of choice for SANs. Fibre Channel is, without doubt, the implementing technology for SANs and made the entire storage architecture movement possible. It is bandwidth efficient and addresses the key problem of scarcity of time that an Internet-driven computing model poses. Internet traffic is reportedly doubling every 100 days. The technology cannot be touched for sustained read-only applications such as streaming audio and video. PC is growing its infrastructure and is working on its high perport cost. It is here, now, and usable.

GE, however, offers backward compatibility and that is a non-trivial advantage. Further, there are now implementations that are sustainable and repeatable. The big issue, perhaps, is the point that any kind of transition takes place. By the end of this year, FC will run at 2Gbps. GE runs at 10Gbps. Broadcom has reported a new chip that will allow 10GB Ethernet to be deployed much sooner than expected. It is possible that the impact of GE may take place as soon as 2001.

Yet none of these transport technologies are standing still and none of them are likely to be cheap. One thing is likely, though. Until the technological positions of these various transports become clear, SANs will continue to lag. Companies need to solve storage problems today and are unlikely to wait while the debates rage. SANS are strategic. Over time, they will provide the best possible means to share storage more efficiently. Yet how will the data travel from point to point?