Friday, October 20, 2006

A burning question for the data center


OK, we have all heard that the data center is going to converge on accelerated Ethernet with Remote Direct Memory Access or on Infiniband that already has RDMA. Clearly, this has not happened and still seems a far away dream given the relatively slow uptake and expense of both technologies at 10Gbits and up.

The real convergence we have seen so far is in a single software stack for both networks under the Open Fabrics Alliance. So--and this is your weekend homework, dear interconnect mavens--what do you suppose the next practical step in this data center convergence will be?

Think on it and keep checking back here for the answer coming in less than a month! --rbm

2 comments:

sink sink socks said...

After which, they would stretch upon that unyielding frame the hide of areality which they had urinsex bilder previously killed.

Anonymous said...

The RDMA specifications within the IETF are final. The IBTA has updated all of its specifications to align with the IETF and these are final. Hardware is being implemented for both specification suites as we speak. So, you ask the question is this all just a matter of software?

In one sense, it has always been just a matter of software. Whether it is OS software, middleware software, application software, management software, it has always been a matter of software. InfiniBand was conceived in 1999 and here we are in late 2006 with still only the basics working and so much further yet to go. Skip past the marketing from the various niche providers and you see a nascent ecosystem that is just starting to come together. The management infrastructure remains largely deficient and unable to plug into the main data center management applications in a meaningful way (yes the plug ins do exist but it is way too manually intensive to be practical on any scale outside some government funded solution).

Over time, OFA and the Unix vendors will provide a credible RDMA ecosystem. Certainly one would hope that IBM's long awaited InfiniBand offering will raise the bar and deliver real value but it is very late to market and IBM remains largely silent, instead hyping Power 6 without filling in the rest of the information gaps.

RDMA itself is under attack. Myricom continues to project that RDMA is unnecessary and the wrong solution as witnessed by their "objective", yet clearly marketing, analysis. Add into the mix companies such as Intel who while they purport to support RDMA such as InfiniBand, their message is very much anti anything protocol off-load. Opening up IOAT is yet another sign that they are working to rally more companies to their banner of processor-centric I/O processing. The problem with all of these critics is they fail to state how they will solve the memory bandwidth consumption problem (can't even agree that there is one) that is only compounded with the increase in the number of cores sharing a given not integrated memory controller.

So, what what is the next step in real convergence? What it has always been? Just a matter of software and certainly more than what OFA produces today.

 
interconnects