IBM wants to be a player in networking, and that means getting 40 Gigabit Ethernet switches out the door to compete with those from Cisco Systems, Mellanox Technologies, Dell, HP, and others.
With the 10GbE ramp well under way at the edge in the data center, thoughts naturally turn toward using faster switches at the aggregation layer – and that's one scenario in which Big Blue's System Networking division expects its RackSwitch G8316 to compete.
There are also customers for which 10GbE switches are not fast enough even at the edge of the network, and the G8316's 40GbE ports will see some uptake here, as well.
The G8316 has 16 ports that run at 40GbE speeds, providing a port-to-port latency of "well below 1 microsecond," IBM claims. The circuits inside the 1U chassis are able to process 960 million packets per second and the switch has 1.28 Tb/sec of non-blocking switch throughput.
IBM doesn't divulge which ASICs it uses inside of its switches, but rival Arista Networks does and it sure looks like a Trident+ ASIC made by Broadcom (not Fulcrum Microsystems, which was eaten by Intel a year ago, as this story originally said).
IBM's RackSwitch G8316 switch
It also looks like the IBM RackSwitch GS8264, which was launched in October 2010 with 48 ports running at 10GbE and four uplinks running at 40GbE is based on the Trident+ ASIC. The important thing is that although it normally takes a couple of chips to build the switch, the Trident+ chip does it all on one piece of silicon, and that like other switch chips, it allows for the bandwidth to be carved up into various port counts and speeds.
The new G8316 switch is, in essence, all uplink – or maybe we need a new word in networking such as sidelink, since it's designed to be used in conjunction with the G8264 edge switches and to handle east-west traffic across server racks as well as to provide aggregation for North-South traffic out to the network backbone and down into the servers.
By using optical breakout cables, which split a single 40GbE QSFP+ port into four 10GE SFP+ ports, you can buy the G8316 today and use it as an aggregation switch, and when you move to 100GbE on the backbone and 40GbE on the edge, you can revert to QSFP+ ports and get a few more years out of the box. And if you don't need 40GbE at the aggregation layer today, you can conversely buy a G8316 and use the splitters to turn it into a 64-port 10GbE switch – and when you need 40GbE at the aggregation layer, you can switch back to QSFP+ ports and do that.
The G8316 supports IBM's Virtual Fabric, which allows for bandwidth to the virtual network interfaces to be dynamically allocated from the switch, and which is particularly useful in virtual server environments where as many as four or eight physical network interfaces are needed in the server to support the bandwidth needs of VMs.
The G8316 also runs VMready Virtual Vision, which is a layer of abstraction that is VM-aware and can switch ports automagically as virtual machines are live-migrated around a network-connected cluster of servers.
Like most switches today, the G8316 supports Fibre Channel over Ethernet (sometimes called Converged Enhanced Ethernet) as well as Data Center Bridging, which is important for running NAS or iSCSI storage protocols atop Ethernet.
IBM will start shipping the RackSwitch G8316 on July 20. It costs $35,999, or just a few pennies under $2,250 per 40GbE port or $564 and change per 10GbE port if you use the splitter cables.
The RackSwitch G8264, by comparison, costs $29,999 and has 48 10GbE ports and four 40GbE uplinks that can also be split up for a total of 16 more ports or the same aggregate 64 ports running at 10GbE. (This stands to reason since the guts of these two switches are essentially the same.) But you are only paying $468 per 10GE port and $1,874 per 40GbE uplink on the G8264, if you do the math.
The flexibility and future proofing you get with the G8316 comes at a price – a 20 per cent premium, to be precise – which is exactly what you would expect. ®