Monday, March 15, 2010

FCoE Isn’t a Replacement for Infiniband, It’s a Cheaper Copy That Customers Will Buy

December 6, 2009 by Greg Ferro · 11 Comments 

Infiniband is some­thing of a for­got­ten pro­tocol these days but many of the mar­ket­ing fea­tures of FCoE are dir­ectly derived from Infiniband con­cepts and architecture.

Paradigms of Data Center Virtualization

For example, this page is taken from the “Paradigms of Data Center Virtualization” by Cisco at Networkers 2005. If you are fol­low­ing Cisco UCS then it should be very famil­iar, per­haps sur­pris­ing the the Unified Fabric idea is not new but either Infiniband or 10GB was going to be the net­work­ing fabric.


Caption Text.

Unified Fabric Basics.(Click for a full size image)

If you are reas­on­ably new to Data Centres then you may not remem­ber that Cisco has a mod­er­ately suc­cess­ful Infiniband product, and in 2005, it was the centrepiece of Cisco’s mar­ket­ing strategy for a Unified Fabric for con­nectiv­ity in the Data Centre. As an aggreg­ator of con­nec­tions from the server down to an Infiniband back­bone it was the Unified Fabric that was going to unite the Data Centre for Grid Computing (the pre­cursor of what is now virtualization).


Caption Text.

Caption Text.(Click for a full size image)

Unified Fabric Server Clusters

And you can see a Case Study of this using a pure Infiniband back­bone for a large scale server cluster in this slide (from the same present­a­tion)


Caption Text.

Caption Text.(Click for a full size image)

So what changed ?

So why didn’t Cisco con­tinue with Infiniband ? I think it’s worth look­ing back and think­ing how we got FCoE instead and look­ing for les­sons to be learned. I can only spec­u­late on what happened.

Recap on Key Infiniband Features

  1. InfiniBand is a high speed, low latency tech­no­logy used to inter­con­nect serv­ers, stor­age and net­works within the datacenter
  2. Standards Based – InfiniBand Trade Association http://​www​.infin​i​bandta​.org and has been work­ing suc­cess­fully for more than ten years.
  3. Scalable Interconnect speeds in mul­tiple of 2.5Gb/s, products are cur­rently ship­ping at 40Gb/​s and 120Gb/​s products have been announced.
  4. Low latency net­work­ing with delays around 20 micro­seconds (end-​​to-​​end) is one thou­sand times less than 20 mil­li­seconds for a data centre eth­er­net network.

I also find it inter­est­ing that the sil­icon from QLogic and Mellanox has rap­idly been repur­posed for FCoE HBA’s. There must be strong sim­il­ar­it­ies is these products for this to happen.

Ethernet is easy to sell

People will buy and use what they know. In that sense, faster eth­er­net or ‘newer’ eth­er­net is an easier decision. I assume that offer­ing cus­tom­ers an easy choice, what they per­ceive as a simple upgrade path, is a much easier sale, than one that needs to teach cus­tom­ers new tech­no­lo­gies. People tend to lazi­ness, and this is a prob­able cause.

Cheap always wins

Ethernet is cheap and not very good at a lot of things but, there is a lot of it about and plenty of exist­ing tech­no­logy. And don’t for­get skills, lots of people under­stand Ethernet and that has a price tag.

And his­tory shows that cheap always wins.

Maybe Cisco wasn’t dominant

Cisco has a policy of being num­ber one or two in any mar­ket. If they can’t do that, they will walk away or do some­thing to kill that mar­ket. For example, in the early days of IPSec VPN, Cisco didn’t have a good story. They bought two or three com­pan­ies before con­ver­ging the code in the Cisco PIX (and later the ASA). They grew to num­ber one by char­ging no license fee for their solu­tion thus block­ing all their com­pet­it­ors from mak­ing a profit. They same sys­tem appears to be hap­pen­ing today with SSL VPN which is now a small license fee for most customers.

If Cisco couldn’t dom­in­ate the Infiniband mar­ket, then per­haps Cisco spun out the Nuova Systems (the com­pany that built FCoE and the Nexus fam­ily sil­icon as a star­tup) to give it a shot at col­lapsing that market.

FCoE cheaper but not better ?

If you spend some time with Infiniband, you real­ise that FCoE isn’t spe­cific­ally a world beat­ing tech­no­logy innov­a­tion. FCoE is the VHS to Infiniband Betamax. And you can see that many of the ideas the FCoE Unified Fabric pro­motes are the same as those pro­moted in the past. In that sense, FCoE isn’t new, just a rehash of old ideas mashed onto exist­ing technologies.

Now, I’m sure that FCoE and Cisco UCS strategy is work­ing just fine for cus­tom­ers and it’s going to do well in the mar­ket. Given that Cisco has spent any­thing up to US$1billion buy­ing, man­u­fac­tur­ing and mar­ket­ing the product, it’s a guar­an­teed suc­cess. But its worth look­ing back on older tech­no­lo­gies with regret and cyn­ic­ally view­ing these new devel­op­ments to determ­ine if they are really the best solu­tions for our networks.

So far, I’m not so sure. We could have had bet­ter, but FCoE will prob­ably work just fine for a while until we need to scale and get faster than Ethernet can ever go. Maybe Ethernet will work for us in the future but Infiniband will still be there wait­ing for us to reuse it. Just like all the other tech­no­lo­gies that we keep reusing.

Please rate this post:

  Why Rate Posts?
1 Star - It\\\'s Crud2 Stars - It\\\'s Tosh3 Stars - Something\\\'s missing4 Stars - Needs works5 Stars - Good Enough6 Stars - Good7 Stars - Excellent8 Stars - Brilliant9 Stars - Astonishing10 Stars - Awesomely Godlike? (3 votes, average: 9.00 out of 10)
Loading ... Loading ...

Comments

11 Responses to “FCoE Isn’t a Replacement for Infiniband, It’s a Cheaper Copy That Customers Will Buy”
  1. Interesting Perspective says:

    I think the most inter­est­ing thing in your art­icle is the cor­rel­a­tion between the two. Great way to bridge the gap.

    Reality is that FCOE is a great sav­ings to com­pan­ies. 10gig is get­ting cheaper but real estate in the DC isn’t. If I can reduce the amount of switches i own but com­bin­ing the FC and Ethernet, great: less power, less cool­ing, less floor space=less $$$$$$.

    For myself as an owner of sev­eral data cen­ters, I want to get more com­put­ing power into every square inch of space, by mer­ging the SAN and LAN into one switch, I can do that. And for that I say ‘Thanks Cisco, now get me nat­ive FC in a Nexus 7k damn it.”

    • Greg Ferro says:

      What’s also inter­est­ing is that the suc­cessor to 10GB eth­er­net isn’t mov­ing very fast. The IEEE is still arguing over whether to have 40Gb/​s or 100Gb/​s Ethernet (which looks noth­ing like Ethernet at all — but that’s another mat­ter) as part of their stand­ards pro­cess. Although early stage test­ing has been done on both the IEEE is still nowhere on standards.

      Compare this with Infiniband at 40GB/​s today, and 120GB/​s around the corner with real, val­id­ated, in use products.

      Sure, today’s serv­ers can’t push 10Gb/​s off the bus, but pretty soon they will be and then some of the bene­fits of converegence will have been lost.

  2. Hi Greg,

    Interesting art­icle. I see you are soften­ing to the fact that FCoE actu­ally does bring some­thing to the mod­ern data centre. Not per­fect by a long shot, but cer­tainly brings some value.

    As for being a rehash of old tech­no­lo­gies and ideas.…. I’m a firm believer that there are no new ideas, just old ideas reborn on faster and cheaper hard­ware. Cycles.

    On a sim­ilar point, I would ven­ture that all tech­no­lo­gies are interim and will even­tu­ally be super­seded. Dare I say even IP will one day be super­seded. But not by FCP ;-)

    Nigel

  3. Brice Goglin says:

    “Low latency net­work­ing with delays around 20 micro­seconds (end-​​​​to-​​​​end) is one thou­sand times less than 20 mil­li­seconds for a data centre eth­er­net network.”

    I don’t know what you did with your Ethernet to get 20 mil­li­seconds, but mine is about 50 micro­seconds and even 5 – 10 if I tune it properly.

    • Greg Ferro says:

      Using ICMP isn’t reli­able at that res­ol­u­tion so you would sup­port that state­ment with methods.

      Also, Infiniband has latency of less than 200 nano­seconds on a single switch, I increased it for a typ­ical con­fig­ur­a­tion with a few adapters and switches. Ethernet can­not come close to that.

      BTW, lots of good inform­a­tion on your com­pan­ies web­site. Must remem­ber to keep check­ing for new information.

      • Kevin says:

        ICMP may not have the res­ol­u­tion, but we get less than 1ms across any por­tion of our DC’s using IxChariot, which is con­sid­er­ably less than 20ms. I agree that Ethernet isn’t as speedy as Infiniband, but you are over­stat­ing the dif­fer­ence by over 2000%.

  4. Adam says:

    As I know Infiniband, is top­ic­ally not data trans­fer pro­tocol, it sends data over IP with MTU size 1492 – 2048 bytes. This is Cisco divi­sion…
    Brocade were invited new pro­tocol FCoE to data trans­fer, Fibre Channel over Ethernet.
    So FCoE is firs imple­men­ted with battle between Brocade and Cisco, first imple­ment­a­tion and ASICs are steel bad.
    And I pub­lish­ing new invents on my site at http://fcoe/ru. Welcome!
    In addi­tional, FC mar­ket is Brocades :-)

Speak Your Mind

Tell us what you're thinking...
and oh, if you want a pic to show with your comment, go get a gravatar!