Friday, September 15, 2006

IBMers have their say

HPCwire has two good articles posted today where IBMers speak out on big systems interconnect issues.

There’s a good interview with David Turek (left) who talks about the Cell/Opteron Roadrunner supercomputer the company announced recently (see my Sept 9 post below). They didn’t ask Dave about IBM’s plans to use coherent HyperTransport in the system, which I think is one of the key details yet to emerge. And Dave said it was too early to discuss the systems software environment, though ironically he took a big swipe at Cray’s heterogeneous HPCS Phase III system as vaporware.

In another posting, Renato Recio (right), an I/O guru at IBM, strikes back at an earlier HPCwire paper from Myricom that took RDMA to task. As I’ve said before (see posting from Aug. 28) I think Ethernet with RDMA has a great opportunity to be a mainstream data center interconnect. What I wonder is why they seem to be taking so long to get market traction.

Know about Roadrunner’s use of coherent HyperTransport? Got some hot insight on the snags with RDMA? Or some skinny on some other interconnect issue? Take a minute to post a comment or drop me an email at -rbm


Anonymous said...

IBM's plan to use Coherent HyperTransport is an important milestone is recognizing the value and impact of having an open, industry standard coherency protocol. It demonstrates the power of such technology and the direct benefit to developers and customers. It further illustrates how AMD leads and Intel follows. Unfortunately, Intel is unlikely to open up its CSI technology or do something bold like move towards industry standards since that would expose its weaknesses and growing inability to compete in industry standard technologies.

Anonymous said...

I agree that Intel is unlikely to ever open up CSI since they are having trouble executing what they are already trying to deliver.

HPC users prefer Opteron because it is the best processor as proven in multiple Top500 offerings. IBM chose HT because it can marry the Cell processor with Opteron much more easily and achieve superior performance compared to other approaches. Can you imagine trying to do this over some non-coherent link like PCI Express? IBM got it right in using HT as anything else would not deliver the performance.

Rick Merritt said...

Hey, Mr and Mrs. Anonymous, let's make it real and share your company affiliation if not your name...especially if you are going to be dising what I assume is your competition!

Anonymous said...

To RDMA or not to RDMA is an old story. Does anyone care what two companies defending their turfs have to say? Customers vote with dollars and that is all that matters. Recio's piece seems more credible so kudo's but is anyone paying attention to this little conflict? Certainly not the press.

Anonymous said...

IBM kept the InfiniBand faith and is banking heavily on its success. They need to replace aging Power and mainframe I/O technology. IBM vision to translate the mainframe experience to the volume space and deliver on the InfiniBand promise will be a challenge to deliver but after 8 years of development, perhaps they have done what others failed to do.

Myricom's position as the once king of cluster interconnects has been waning of late. Be foolish to count them out given their shift in focus to 10 Gbps Ethernet switches as shown at Hot Interconnects and other venues the past couple of years.

The critique was flawed on various aspects. That is easy to see by anyone who comprehends interconnects. But that does not really matter. The fact is RDMA or any new technology requires a complete or near-complete understanding of the solution apace. This makes it hard for anything new to succeed and is at the heart of why so many new technologies fail.

So, hat's off to IBM or any who buck the industry wisdom and for making a go of it. It is rare to see such a large company take such a major gamble when all others turned away and took a different road to travel.