Why storage over IP?

Posted on September 01, 2000

STEVE DUPLESSIE -- The Enterprise Storage Group
Click here to enlarge image

One year ago, the thought of storage over IP seemed ridiculous, and few took the concept seriously. However, a number of vendors were working in stealth mode on a number of technologies that might make storage over IP feasible. In a way, those developments culminated this summer when Cisco tipped the apple cart with its acquisition of NuSpeed (see p. 1). These actions give rise to a number of interesting questions.

Why would we need/want IP for our storage network?

Because (1) we already have IP everywhere and (2) independent storage networking is hard, complex, and expensive. If we could only have one network, it would be much easier.

IP is inefficient for storage, and performance is terrible, isn't it?

No, TCP is inefficient for moving storage data around on IP networks. But that may be fixed with technologies such as hardware NIC accelerators, and by vendors such as Giganet (see p. 1) with technologies such as VI. Other vendors are using their own transports instead of TCP, which will start a religious war.

I recently saw numbers from Giganet claiming that it was transferring data at full Gigabit Ethernet rates, while consuming less than 8% of the CPU cycles. Think about those ratios with 10Gigabit Ethernet and the story becomes really compelling.

Don't we move only file-level data over a network, and block data directly?

Today, we move files around a network using CIFS or NFS, through the file system and the full kernel stack. By the way, the Virtual Interface (VI) architecture will let us bypass 90% of those kernel operations, which is why everyone is so excited about it.

We don't move "block" data around IP networks yet. An application calls a SCSI device driver as its purest, rawest form of data movement. It calls for blocks of space on a disk. Block transfers are very fast. The biggest trick to using IP for a storage network is going to be the ability to do logical SCSI block calls over an IP wire. That's one big reason Cisco bought NuSpeed-they figured that part out. They aren't the only ones working on this problem, however, and you can expect a number of solutions over the next nine months or so.

What's going on with iSCSI?

iSCSI is the proposed IBM/Cisco standard to move block data over IP networks. (See InfoStor, June 2000, p. 1.) Nishan has its own way of doing this, called Storage over IP (SoIP), but I think that the odds are that some form of iSCSI will win out. I predict that Nishan will try to get SoIP incorporated inside iSCSI, or to get support for both.

What do the disk drive and disk array manufacturers think about all this?

For the most part, they don't really care. NuSpeed's claim to fame is the ability to take a Fibre Channel port in front of a disk array and turn it into an IP-enabled SAN device. In other words, a server does a BLOCK data call via the SCSI driver (the same way it always has), but the transport used is IP via a Gigabit Ethernet card, which connects to the NuSpeed box, which in turn connects to the Fibre Channel disk array.

If anything, the disk vendors love this because suddenly SANs become infinitely easier to configure, support, grow, etc. Current SAN users complain about the complexities, lack of interoperability, cost, etc.

For more information e-mail gregd@enterprisestoragegroup.com, and request the full (free) Storage over IP report, which was published last month.