Monday, November 05, 2007

Give me iWarp speed, Scotty

Infiniband has a pretty comfortable position as the growing king of cluster interconnects right now, according to Bob Wheeler of market watcher The Linley Group. Gbit Ethernet has peaked in the Top 500 list of supercomputers, IB is stealing market share from Myricom and Quadrics and the 10G Ethernet bandwagon is bogged down in the mud, according to Wheeler.

He notes that neither Dell, HP nor IBM have started using iWarp chips from the likes of Chelsio and Net Effect in their mainstream servers yet. "Until the big OEMs start investing in the iWarp ecosystem, it is not going anywhere," Wheeler said.

I think I wrote my first story detailing the folks pulling together the iWarp standard back in 2001 or so. HP and other major OEMs were among the big backers then, but today HP is collaborating with Netxen, a 10GE vendor that won't support iWarp until next generation products come out.

Making matters worse for the 10GE crowd, Myricom is rolling out products using its own protocol over Ethernet, and Mellanox plans to support Infiniband on top of Ethernet as an alternative to TCP. "They are fragmenting an already small market," Wheeler said.

The flip side of the coin is that IB is not likely to extend far beyond high-end clustering apps, Ethernet will dominate LAN connections and 10GE will eventually go mainstream once 10GBase-T prices come down in a year or two. Chip designer Fulcrum is giving 10GE a little nudge forward with some new products rolled out today.

"Based on the numbers I've seen from systems integrators, it seems like IB has a significant price/performance advantage over 10GE now, but that will change rapidly as 10G E ramps into volume," Wheeler says.

So, at the end of the day, the IB folks may be crying "Beam me up, Scotty!"

3 comments:

Anonymous said...

Does anyone truly believe analysts these days? Sure, analysts can cite statistics, polling data, and so forth but as anyone knows, anyone can make statistics and polls support nearly anything they desire (just look at US politics for daily proof that there are lies, damn lies, and statistics). Recall that major analyst who said 80% of servers would be IB enabled over and over again perhaps because Dell, HP, and IBM were co-creators of IB. Never could pin that person down as to what exactly "IB enabled" meant since it seemed to be a sufficiently fluid term to be largely useless. No matter its meaning, the analyst was so obviously grossly optimistic that the 80% number was quietly replaced with a new analyst spin on how IB would be positioned and adopted. Funny, if the analysts had actually asked the server vendors what they were doing, perhaps they would have avoided this entire marketing fiasco and provided some real fact-based analysis.

Fast forward a handful of years and one analyst has learned from the past mistakes associated with over-hyping a technology but now has swung the pendulum the other way with a very negative view of iWARP adoption because Dell, HP, and IBM are not shipping iWARP devices in their servers today. What companies are cited....a handful of small basically start ups who constitute a very small percentage of the 10 Gb Ethernet market. Unclear from your blog whether the analyst discussed Intel or Broadcom product offerings who largely dominate the server Ethernet market - oh, and let's not discuss Broadcom's recently announced iWARP 10 Gb Ethernet products which could provide a natural evolution of plain 10 Gb Ethernet to higher volume iWARP attached servers.

I'm unwilling to pay thousands of dollars for some analyst report to comprehend how the 10 Gb Ethernet market is going to evolve when it is obvious to anyone within the industry. Perhaps it is covered there but what gates iWARP adoption can largely be laid at the feet of 10 Gb Ethernet - both vendors and technology.
10 Gbps optics remain too expensive and over priced.
10 Gbps CX4 (also used by IB) has a number of mechanical problems that complicate deployment. Active cables help out some but are not a panacea given the CX4 connector and their costs.
10GBase-T spec is very new so very few if any products have come to market yet.
10GBase-T implementations are power pigs making it impractical to deploy in servers for anything more than 30m.
10 Gb Ethernet switch offerings are way too expensive and over priced.

All of these factors have contributed to the less than stellar adoption of 10 Gb Ethernet. Oh, and if the analyst pay attention they can find nearly all of these reasons cited in the various marketing materials from the IB vendors as to why IB is superior to Ethernet.

So, here is a 10 Gb Ethernet adoption quagmire that obviously weighs heavily on iWARP adoption - can't adopt a new technology if the interconnect isn't cost effective or available to deploy in any significant volume today. Perhaps this hidden in that expensive report but what is publicly cited here is iWARP is going to happen any time soon because three server vendors have yet to ship iWARP devices from a couple of very small start ups. Let's not talk about the hardware and business aspects surrounding 10 Gb Ethernet adoption itself and focus instead of what three server vendors are doing. Let's not talk about how IB has come into its own over the past seven years largely independent of these same server vendors who obviously did not buy into the 80% IB enabled analyst hype. Let's not talk about the other aspects that gated IB adoption and to date have gated iWARP adoption - software is just as important as the iWARP and 10 Gb Ethernet hardware.

I will agree with this analyst that IB isn't going to take over the world as the prior analyst proclaimed. IB's penetration within the data center will be gated by many factors but not because the IB transport is running over Ethernet. It is much simpler than that - it has to do where IT spends its budget and how good enough technology will evolve and be used to supplant existing excellent technology. That is a much longer response and I've spent too much time today writing this one as it is.

To be clear, I have nothing against analysts - off times there are insightful observation here and there in their reports. However, many of these observations really come from the technology and business leads they interview to produce their reports and not from examining the attach rates of start vendor products.

Anonymous said...

yeah, all the same could have been said about Ethernet when it was 10base2-T when 10baseT or 10/100 was on the horizon.

10Gbase-T will happen just like every other copper Ethernet standard at distances making data center deployment totally practical. Why? Because there's good money in it, based on the last 10 successful turns of the crank.

In 5-10 years, 10GbaseT will come to contain Infiniband to the tiny niche HPC segment.

Computer Consulting Kit said...

It seems to me that because of the plethora of companies releasing technologies before they are fully developed in order to keep up with the competition, there is a lot more hype than there used to be about new technologies, whether from the companies themselves or from analysts. I think while new technologies might be very exciting, it pays to wait until they have been a bit more developed (and until some of the kinks have been worked out) before counting on them as the next best thing. I talk a lot about “early adoption” and other topics for businesses in my blog

 
interconnects