Is iSCSI really dead? The Register thinks so ...
The Register thinks that FibreChanneloverEthernet(FCoE) will take over the datacenter and proclaims “iSCSI: Game over”. Okay, no problem …. there is already a FCoE implementation for Solaris. Been there - done that. There is only one simple problem: I´m not sure, that FCoE is the way to go.
At first: You don´t get really an advantage by FCoE in relation to iSCSI. Using FCoE for storage connectivity means: SCSI commands over Fibrechannel over Ethernet. iSCSI is SCSI commands over IP over Ethernet. Same number of layers. So nothing to win here. And it doesn´t look as way to a single fabric datacenter. I don´t think that even a more evolved and specialized Ethernet can solve the problems of increasing latencies when you use it for several services at the same time. What´s more important writing the data to the client or getting the data from the disk? But let´s assume, the latency issues are solved in an acceptable manner. In return of using FCoE, you loose proven technologies: You can´t route between networks, you can´t use IPsec for encryption and authentication and integrity checking. In IP you have already a proven wide-area naming service (you call it DNS) and you have your IP networks available today. In iSCSI you have alreay an established naming system calles iSNS. With hexa- and octaicore processors you have more than enough clock cycles for the TCP/IP stack. There is an important fact: The E in FCoE isn´t the Ethernet of today. You can´t simply use your existing Ethernet-Network. FCoE needs an enhanced Ethernet. It isn´t usable on normal Ethernet as Ethernet is a lossy transport by nature. Higher protocols in the TCP/IP stack solve this problem for the application at the moment. But the Ethernet of FCoE has to do this on its own. So you need switches capable of providing a lossless Ethernet to use them as your storage connect. iSCSI simply uses TCP/IP to provide a lossless communication, thus you could even use a wide area network for a storage interconnects. Try this with FCoE) .The TCP/IP stack and adjactent services (routing,naming,encryption,authentication) are a proven piece of software in most operating systems with a common set of standardized protocols with their share of time to iron out incompatibilities. Why should we develop and debug all this stuff again? The whole FCoE talk looks like the last stand for the FC vendors to fight for their market share before storage networks are migrated to IP as well. And when you really want to establish a decent new storage interconnect: Use Infiniband as the foundation. 40 GBit/s (so you have more than enough headroom for multiple services in parallel) and really low latencies. It gives you iSCSI as a well established standard and RDMA over Infiniband to take TCP/IP out of the equation.