Fulcrum Microsystems Blog

By Gary Lee

The market for cloud computing has taken an interesting turn with Amazon’s announcement last month of its Cluster Compute Instances service.  The announcement makes Amazon one of the most high-profile firms to deliver cloud-based HPC service and marks the growth phase of the market for cloud services that depend on high-performance, low-latency networking to deliver their value add.

As reported by Network World, Amazon said its new service provides the same performance as a custom-built infrastructure with pay-as-you-go pricing.  The company said its cluster service delivers up to 10 times the network throughput of its existing EC2 cloud computing services.

As it turns out, many large companies were using the EC2 service for computing jobs that they would normally run on an HPC infrastructure. To respond to this trend, Amazon didn’t change out the computing infrastructure, but rather boosted the performance of its cluster networking gear.

We think many more cloud computing vendors will also make the move to a low latency infrastructure to both keep pace with Amazon and others, but also to help improve their profitability.   If a move to a new infrastructure means a 10X increase in throughput – as Amazon has announced – then that means also that the company can schedule more customer jobs on the same server infrastructure.

Cloud computing performance used to be measured in terms of bandwidth in and out of the datacenter.  That won’t change for some folks, like Facebook, but those companies that sell actual computing cycles will need to consider the impact that a low-latency infrastructure will have on their services.

Comments are welcome: feedback@fulcrummicro.com.

By: Gary Lee

Normally, we limit the focus of this blog to lossless fabrics for telecom and data centers.  But those worlds are ultimately impacted by bandwidth consumed by users like you and me.  So I thought this recent article in Wired was worthy of comment. It discusses how the bandwidth offered by Verizon’s FiOS fiber data services for consumers has been steadily climbing and is now up to 10Gbps to households in some communities.

With HD televisions and triple-play network services requiring more bandwidth per household, service providers are looking at ways to deliver that bandwidth – all the way to today’s current maximum standard of 10-Gbps. Almost all of these networks are becoming packet-based, with Ethernet the dominate layer 2 protocol.

This skyrocketing demand for Ethernet-based data services is not going away, so how long will it be before carriers are looking to even higher data rates?  Look for 40G and then 100G coming to a set top box near you!

Comments are welcome: feedback@fulcrummicro.com.

By Gary Lee

There has been a lot of buzz lately around the potential of the OpenFlow Switching Protocol to be the control plane solution for the data center interconnect.  For those not familiar with OpenFlow, here’s how I described it in a previous post:

OpenFlow is described on its Web site as an open standard that allows researchers to run experimental protocols on existing networks.  The effort began at Stanford in 2008 and has installations at a variety of college campuses across the country.

The latest excitement is mainly due to the software’s potential for reducing both hardware and software complexity, as well as the potential power savings that comes with central control of resource allocation. Managers of large data centers want hardware vendor flexibility along with a simplified — yet extensible — software stack. OpenFlow has the potential to bring them both.

At the same time, the IEEE is defining DCBx as a way for networking equipment to automatically exchange capabilities in order to create a unified data center network configuration, based on common-denominator functionality.

These overlapping approaches provoke the question of which is a better way to configure a large data center fabric.

DCBx appears to be leading the race with several software implementations now available. In addition, OpenFlow controllers will need to work with legacy data center fabric technology in the near term, including DCBx enabled fabric domains, and are therefore focusing on provisioning resources at the network edge. But in the long-term, OpenFlow could provide a central means of controlling and provisioning the entire data center fabric without the need for DCBx. Should be interesting to see how this plays out.

Comments are welcome: feedback@fulcrummicro.com.

By: Gary Lee

There’s a lot of buzz about 100G Ethernet, with customers saying they need it to solve pressing bandwidth shortages in the carrier network and in the data center.  Some vendors are also caught up in the frenzy and are claiming they can deliver it now.

But a recent press release from the IEEE should serve to remind the market that there’s still work to do to make 100G viable.  In fact, the headline from FierceTelecom sums it up nicely: IEEE’s 802.3ba standard sets stage for 40, 100 Gbps Ethernet – with an emphasis on the stage being set.

We predict that in the enterprise and data center, 40G will emerge in a significant way before 100G due to the complexity and cost of designing the switch and NIC chips.  These chips need to aggregate 10 lanes of 10GbE to deliver 100G performance; that’s a complex design that requires costly cable assemblies.  It looks like 4-lanes will be the configuration of choice for the near future as is used today with 40G.  What’s needed is 25G SerDes that are cost effective, but this will require 28nm processes, with production products still years away.

Comments are welcome: feedback@fulcrummicro.com.

By: Gary Lee

In another sign that Ethernet is taking over in many parts of the telecom network, here’s an article from Light Reading.com about how T1s needed for backhaul networks may die as soon as next year, being replaced by Ethernet:

“Wireless network operators are requesting fewer T1s and more Carrier Ethernet connections for mobile backhaul, and could stop ordering T1s entirely as soon as next year, Tower Cloud Inc. CEO Ron Mudry told attendees at the Backhaul Strategies and Core Convergence for Mobile Operators event here today.”

The article goes on to detail the various costs associated with building carrier networks, including fiber trenching, electronics, rent for cell sites, etc.  With the sum of these costs, though, carriers need to be judicious with the types of networks they deploy.

While they can’t always future-proof their networks, carriers must anticipate increases in bandwidth demands from advanced mobile applications and look to implement scalable network solutions that offer clear upgrade paths to 10G, 40G and, eventually, 100G data rates. They must also do this while minimizing costs.

With new and advanced congestion management, QoS and network clock synchronization capabilities, Ethernet will be the packet backhaul technology of the future.

Comments are welcome: feedback@fulcrummicro.com.

By: Gary Lee

One of the major technological advantages our FocalPoint switch family enjoys over competing 10GbE switch chips is its high-performance circuit technology based on asynchronous logic.  Among the primary features this technology enables is a shared memory output queued architecture that provides Fulcrum’s industry-lowest latency and excellent multicast efficiency.  Lower power is another benefit as only the circuits that are working draw power, in comparison to a clocked chip where the constant pulse of the clock causes all circuits to draw power regardless of whether they are working.

Asynchronous technology has been around for a while and is now starting to reach a critical tipping point in the industry.  A number of reasons for the slow roll out exist, but two of the primary ones are that asynch designs are very labor-intensive, and it’s been difficult to get synchronous and asynchronous components to play together on one chip.

This article in EE Times from Alban d’Halluin, director of product marketing at Tiempo, discusses the issues associated with asynchronous chip design and the way to help it to move to mass adoption—develop better tools and flows.

As asynch pioneers, Fulcrum has developed its own tools and methodologies and now have four generations of commercial chip designs under our belt, and are working on a fifth.  So we know how to get this done.

However, the emergence of tools vendors like Tiempo is exciting for the industry.  Certainly asynchronous design goes against the comfort zone of 50 years of clocked chip design, and some competitors have sought to use that fact to keep customers from realizing the benefits of clockless chips.

Better commercial design tools shows a maturing of the industry that will inevitably make more folks comfortable enough to see the benefits that the technology brings to a wide variety of devices.

By Gary Lee

I wrote a feature for this month’s issue of Lightwave magazine in which I discussed the details of switch chip implementation for carrier class optical transport systems (OTS).  Ethernet-based switch backplanes require different considerations at the chip level.  Here’s an excerpt from that article:

Carrier-class optical transport systems (OTS) have a backplane fabric based on a legacy of carrying TDM traffic that has been maintained in some form even into this age of IP-oriented data traffic. As we move toward a completely packet-based world, traditional TDM-based OTS are evolving to more efficient backplane architectures using high-bandwidth Ethernet switch chips, where TDM data is transported across a packet fabric.

Designers, then, must know how OTS requirements map into Ethernet switch chip features. This includes interface protocols, port count, traffic classification, congestion management, and how to choose from the existing traffic tunneling methods that include the use of proprietary headers to identify unique traffic flows.

These switch chips must support a wide range of features while also minimizing system costs. On the customer side, a variety of interfaces and protocols must be supported including Ethernet and storage, such as Fibre Channel or Fibre Channel over Ethernet. The customer traffic must be efficiently aggregated for transport over the optical carrier network while maintaining service-level guarantees. On the carrier network side, the switch chips must provide high-bandwidth connections for Optical Transport Network (OTN) or SONET/SDH transport.

Traditional OTS backplanes have used cell-based fabrics where a limited amount of packet traffic was segmented into cells on ingress and re-assembled into packets on the egress.  If the OTS includes Ethernet aggregation capabilities, it must combine multiple lower-rate ports into high-bandwidth links that support optical transport protocols.

Read the full article here.

By: Gary Lee

In a move that’s bound to compete with data center giants Cisco and Juniper, Brocade has announced Brocade One, its converged data center architecture.  Per this Network World article, the solution is virtualization-heavy, aiming to transform the data center from physical assets into virtual services that can be allocated via software commands.

Virtualization is a topic that we’ve covered before in this blog, but this recent announcement from Brocade brings a new look to the data center.  The same data center bridging (DCB) and scalable virtual endpoint support that make FocalPoint silicon an ideal choice for flattening the data center architecture also make it an ideal switch candidate for a virtualized environment.

FocalPoint delivers low latency along with the scalable architecture and congestion management features necessary for lossless data transmissions and enables fabric solution providers to deliver innovative platforms that are the foundation for some of the largest, highest-performance data center fabrics in existence.

By Gary Lee

On Wall St., it seems like latency is the only thing that matters in the network these days.  That’s because brokerages are using algorithmic trading programs in a race to make trades – with the first one to execute gaining a price advantage.  Often times, it’s the latency in the network that makes the difference between which company makes the trade first.

That makes Wall Street tech tradeshows a great venue for announcing new low latency standards, and the InfiniBand Trade Association took advantage of the HPC Financial Markets show last week to announce RoCEE – the RDMA over Converged Enhanced Ethernet standard.  Remote direct memory access (RDMA) reduces endpoint latency by transferring data directly from the memory of one computer straight to the memory of another without being slowed by the operating system or the NIC memory.  RoCEE (which the IBTA is calling “RoCE”) brings RDMA to converged Ethernet architectures, much like the Internet Wide Area RDMA Protocol (iWARP) did for TCP.

RoCEE will be an interesting development for traditional InfiniBand fabrics that can now benefit from the cost-effectiveness of Ethernet, and, using FocalPoint, a very competitive fabric latency as well.  According to press reports, RoCEE should bring endpoint latency down from about 4.5 us that is standard with Ethernet to about 1.3 us.  RoCEE layers InfiniBand’s layer 2 / 3 protocols on top of Ethernet’s physical and MAC layers.

What will be the impact on the market?  Timothy Prickett Morgan of The Register asks point blank whether RoCEE will be the death of InfiniBand, before concluding probably not, but it looks like an “octogenarian’s birthday party.”  Naturally, the IBTA disagreed with that analysis.

There will always exist a need for the highest levels of performance that are now served by proprietary fabrics at the very high end, followed by InfiniBand which currently offers 40G of link bandwidth with sub microsecond endpoint latency.  But, Ethernet with RoCEE will now be able to take some of that market, where 10G of link bandwidth with low latency fabrics and endpoints can provide a significant improvement over traditional RDMA solutions.

RoCEE products should be announced soon, and it will be interesting to test them along with a FocalPoint fabric to show this improved performance.  In the end, though, we believe that RoCEE signals the continued expansion of Ethernet in the data center and a graceful way to leverage emerging Ethernet technologies.

By: Gary Lee

Network security was a central focus at the Linley Tech Spring Conference last week, where I presented on advanced memory architectures to an audience of about 60 engineers.

The network security theme resonated with me due to several key features that make FocalPoint ideally suited for enabling network security applications.  In fact, FocalPoint has gained strong traction in this market due to features such as an advanced header classification engine that can preprocess packets and insert the result of the classification in a header tag.  This feature can help reduce the load on down stream processors, and the classification engine can also direct flows to the proper processing resource (CPU or NPU) based on deep header inspection.

Many of the Linley network security presentations brought up the problem of dealing with high bandwidth 40G and 100G network connections. FocalPoint’s advanced hashing algorithms allow efficient flow distribution across multiple security blades in a chassis switch, effectively reducing the bandwidth load on each blade.

FocalPoint also provides symmetrical hashing and the hash key can be derived for various user defined header fields. Simple fail-over mechanisms allow flow re-distribution in the event of a security blade failure. Also, the low cut-through latency of FocalPoint makes security processing a true “bump in the wire,” with very little impact on network performance.

Overall, it was a well-attended event, and I received many good questions after my presentation.  Click here for a copy of my slides.