Moving TCP/IP Into Hardware

Both companies have engineered the technology to address a major IP storage hurdle. They've developed accelerator cards that put the TCP/IP processing into hardware for reliable data transfer across the network, a significant, necessary step if IP storage hopes to compete with Fibre Channel's performance levels.

But this is where the similarity between the two companies ends. Each uses a different method to design the chips on the accelerator cards. Alacritech, a San Jose, Calif.-based provider of TCP/IP acceleration solutions, has approached the technology from a networking viewpoint by developing an accelerator card (also known as a data path off-load adapter) that does a partial TCP off-load. That means the TCP data transfers into hardware, while conducting the connection process via software.

However, storage-based Adaptec, Milpitas, Calif., has designed a product from,fittingly,a storage point of view. Its Storage Protocol Accelerator and Storage Router Accelerator, announced last quarter, have customized silicon that does a full TCP off-load, meaning the entire TCP/IP process takes place in hardware.

"The host performance is absolutely minimal," says Ram Jayam, Adaptec's vice president and general manager of the storage networking group. "We will not add more overhead to the host. If we had to involve the host to do TCP implementations, then there would be a bottleneck on the host performance."

Clearly, Alacritech has different reasoning.

"I say, 'Who cares?'" says Joe Gervais, director of product marketing. "If you do it that way, you break fundamental things you don't want to break. You won't be able to take advantage of the standard Ethernet protocol for load balancing and failover."

All of that is part of the overall challenge in making IP storage a competitive technology against Fibre Channel, according to industry experts. Fibre Channel was designed specifically to handle storage-block data with high reliability standards; there is a tight handshake when data is sent between devices. IP, however, was developed as a "best-effort-to-deliver-data" technology. That is not acceptable to storage administrators in charge of mission-critical data.

"When we do TCP in an ASIC [application-specific integrated circuit, we can give the same performance that Fibre Channel is giving now," Jayam says. "And we do reliable data transfers with low CPU utilization. The host CPU does not get involved in retrying packets. This is completely acceptable in the storage industry."

But Gervais argues that implementing the connection work via software is optimal because connecting is generally not a frequent task, and placing the actual data transfers in silicon makes more sense because that "is where the heavy lifting is done."

"You get huge advantages in failover and load- balancing by leaving [the connection in software," Gervais says. "By doing a full off-load, you are putting the TCP stack in a room that can't talk to other TCP stacks. You have two autonomous entities not talking to each other. TCP is a connection-oriented protocol. It needs a central coordinator, and the best way to coordinate is on the main system CPU."

Besides, he says, storage vendors are coming into an

Ethernet world.

"Network administrators expect to use Ethernet methods that they have been using for years," Gervais says. "These two worlds are colliding. This is not just a storage play and companies like Adaptec ...don't realize that."