As published at www.openfabrics.org
I learned three important things about OFED at the recent OFA workshop in Monterey, CA:
1) OFED is critical to a significant and growing number of organizations, but …
2) Almost none of those organizations realize it, and that is why …
3) The perception of OFED is a significant limiting factor to several large growth opportunities
Because of its heritage with the OpenFabrics Alliance, OFED is primarily perceived as an InfiniBand-only technology, despite the fact that it can be
applied just as readily to Ethernet fabrics. Furthermore because of its distribution methods, it is reasonable to assume that most InfiniBand users
have OFED in their software stacks. That’s about half of all HPC clusters. OFED should be fairly well penetrated into HPC.
But our surveys would indicate differently. In two years of comprehensive HPC software studies, only 1% of respondents identified OFED as part of their
software environments. Conclusion: most people are unaware they’re using it, or if they are, it’s not important enough to make it onto their mental
list of software tools.
As OFA seeks to grow the penetration of OFED, I therefore propose the following question:
Is it enough to get more people using OFED,
or do you also need them to know they’re using it?
This is not a rhetorical question, and it is one that OFA should discuss seriously, as the answer can dictate different courses of action. If unconscious
usage is sufficient, then under-the-covers distribution partnerships can increase share. If conscious choice is required, then OFA must consider
how to brand OFED in a way that will allow it to expand beyond its current HPC InfiniBand cluster niche.
Because of its current IB association, OFED is likely missing the half the HPC cluster market that is running on Ethernet. Furthermore InfiniBand is
not well penetrated as a storage interconnect, and even less so as a LAN, whereas one-third of HPC Ethernet users have it as all three. An Ethernet
embrace will also be necessary in order to expand to High Performance Business Computing (HPBC) applications in areas like financial services,
retail, and logistics, which are largely Ethernet-oriented, and it is a must for the much larger non-HPC datacenter market over the horizon.
Along the way, OFA will also need to consider the requirements beyond performance in business computing environments, where end users are less likely
to take risks on new technologies. (In HPC, once the problem is solved, it’s no longer interesting, and you start the next problem. In enterprise
datacenters, once the problem is solved, for heaven’s sake, don’t touch anything ever again.)
At this year’s conference, it was evident that there are tremendous potential growth opportunities available for OFED, and it is also clear that there
is a dedicated group of individuals ready to tackle the challenges. At the top of the list is a superhero-or-mild-mannered-worker identity question
for OFED. Being Superman is nice, but business-oriented datacenters may be more likely to employ Clark Kent.