Cisco Data Center Fabric -- An Overview- Cisco VP Soni Jiandani

Silvano Gai: Nexus UCS

Silvano Gai: Nexus UCS

Fulcrum's Formula

Fulcrum Microsystems is proud of its FM4000 family of switches which come in a variety of port configurations and package sizes that are tailored for key applications ranging from server clustering to port aggregation. But I am interested in talking about its FM4224 model which provides ultra low latency switching.  To understand how Fulcrum does this we need to understand the two very common Ethernet switching paradigms first: cut-through and store-and-forward.  

Both Store-and-forward and Cut-through switches follow the same forwarding decision process which is based on the destination MAC address of data packets.  It is the series of steps in which they do this forwarding decision that differentiates the both processes.  Ideally store-and-forward switch makes a forwarding decision on a data packet after it has received the whole frame and checked its integrity, a cut-through switch engages in the forwarding process soon after it has examined the destination MAC (DMAC) address of an incoming frame. The DMAC address is found in the first 6 bytes of a frame so as per above definition a cut-through switch should screen only those first 6 bytes but in practise they wait until a few more bytes of the frame have been checked before they make any decision.

Let’s look at cut-through switching in more detail.  When a cut-through switch receives and invalid or bad packet it does not drop the packet like a store-and-forward switch but the best it can do is flag it and send it. So all the invalid packets are sent to the other segments of the network where the frame is examined and dropped by the destination switch.  So to make a more intelligent decision cut-through switches examine more than just the DMAC bytes when a packet comes in. And the more interesting thing is that the number of bytes the switch is going to read can be predetermined based on the value of the EtherType field. For more information on the EtherType field of a frame please visit http://en.wikipedia.org/wiki/EtherType .  

So what’s the major advantage of the cut-through switches over store-and-forward switches? Well you have guessed it right. It’s the speed at which they forward the packets. And for some high performance applications where latencies are very critical cut-through switches dominate over the store-and-forward switches. But cut-through switches are not all that great. They have some drawbacks too. Check out the link which shows a presentation from Fulcrum’s website. The presentation compares different switching paradigms and their drawbacks.The comparision is based on the memory architechture.

But my question is what is fulcrum's formula? If you checked the above link you will see that Fulcrum uses RapidArray memory and Nexus crossbar to improve it buffer speed. So Nexus is a highly efficient fully-connected non-blocking crossbar circuit, providing the highest capacity of over a Terabit in TSMC’s 130nm FSG process technology, the lowest latency of less than 3ns, the smallest die area and the most efficient power profile where power consumption is directly related to activity. Nexus has been fabricated and fully characterized in process technologies ranging from 0.25um to 65nm. It is being leveraged by Fulcrum in its latest generation switch products, and licensed to partners as an efficient on-chip interconnect infrastructure.

Fulcrum has designed its own SRAM and TCAM technology, to deliver sophisticated multi-ported SRAM blocks that operate at more than twice the speed of standard ASIC-based memories, and consume less power (only when active). The circuit complements Nexus in switch systems where fast memory and efficient switching go hand-in-hand. Similar to Nexus, RapidArray has been fabricated and fully characterized in various process nodes and block sizes, the largest to date being well over 9MB of SRAM and 500KB of TCAM in the TSMC 65nm process.

Together as shown in the above diagram Fulcrum provides single shared memory for highly efficient cut-through switching.