The OpenFabrics Alliance develops, tests, licenses, supports and distributes open source software for the HPC interconnect world with a focus on low latency and massive scalability. We caught up with Paul Grun, Vice-chair of the OFA and Co-chair of the OpenFabrics Interface Workgroup and Jim Ryan, Chairman of the OFA to discuss the 1oth annual gathering of the group. The 2014 Annual OpenFabrics International Developer Workshop will take place March 30 – April 2 in Monterey, CA.
insideHPC: Congratulations on achieving your 10th annual OFA Developer Workshop. Can you tell us a little bit about the OpenFabrics Alliance and the workshops?
OFA: The OFA’s roots go back to the emergence of the InfiniBand architecture in early 2000. It began as an effort to prevent the fracturing of the nascent high performance interconnect industry. The idea was to come up with a single set of APIs, libraries and other software components to support modern interconnect technologies, and to make them available through open source. Accomplishing that required an unusually high degree of collaboration among the hardware and OS vendors, software developers and end users – the kinds of collaboration you can only get through an open source community. The result was what we know today as OpenFabrics Software (OFS), which is available for both Linux and Windows. Over the years the alliance’s charter has expanded beyond the bounds of the InfiniBand Architecture so today OFS runs over a variety of industry standard networks including InfiniBand, RoCE, iWARP and a few others. The annual workshop is the most publicly visible representation of the open source community that powers the OFA
insideHPC: What is the significance of these workshops?
OFA: Over the years the OpenFabrics Alliance has developed a record of technology leadership, and you only get that kind of leadership thinking through the sort of face-to-face collaboration that occurs at the annual workshops. These workshops are where I/O technologists from all branches of the computing world can gather in an environment that fosters creative solutions to problems such as those posed by emerging Exascale computing. The outcome of these workshops ends up being a key driver in focusing the OFA’s efforts in continuing to push I/O technology forward.
insideHPC: Can you give us an example of the sorts of outcomes that typically emerge from these workshops?
OFA: At last year’s workshop the notion of “application centric” design emerged, which is the idea that application requirements should be the key driver in defining the interface that an application uses to access the underlying network. Traditionally, interconnect and protocol designers produced a network architecture from the bottom up and then left it to application developers to code to the resulting APIs. At last year’s workshop, it was noticed that this sometimes produces APIs that are not well matched to the applications that use them. This, in turn, produced the suggestion at the workshop that we turn this paradigm on its head by driving network stack development from the application level down. The OFA took that input to heart and created the OpenFabrics Interfaces working group (OFI WG) which has spent the past nine months applying these application centric principles to the definition of the next generation network software stack. This new work group is chartered with making sure that the next generation of OFS meets the needs not only of HPC applications where the OFA has traditionally been a leader, but also several other broad classes of applications outside of HPC. The work group now has 20-25 active participants at its weekly meetings, with the majority being members of the OFA but also including contributors from the industry outside of the OFA as well.
insideHPC: What is influencing this year’s agenda?
OFA: Three things, really. One is this focus on application-centric I/O and all that that implies, another is the on-going emergence of some important technologies such as Non-volatile memory, and the third is the prospect of Exascale computing just over the nearest horizon. Since all three of these are disruptive, the theme for this year’s workshop is “Disruptive Technologies”.
For example, we already mentioned the new OpenFabrics Interfaces working group, whose work has the potential to be highly disruptive (but in a very positive way!) This is a “disruption” that the OFA is encouraging because we think it has the potential to help us deliver interconnect technologies that substantially improve application efficiency.
Beyond this ‘disruptive software technology’, there is a continuous stream of emerging hardware technologies any of which can have a profound effect on the role that I/O plays in classical computer systems. Technologies such as NVM that we mentioned above, new processor architectures, evolving physical layer transports and so on.
And to help us address the looming issues of Exascale computing, we are planning several sessions to discuss the extreme requirements of modern radio astronomy.
All of these sessions are designed to stimulate conversations on how best to improve application efficiency through leading edge interconnects.
insideHPC: What sort of outcomes would you expect from this year’s workshop?
OFA: Given the open-ended nature of the workshops, we’re often surprised at the results that emerge from year to year. However, based on the results of last year’s workshop and the steps that the OFA has taken since then, we expect that one outcome will be a coalescing around the work now being done in the OFI WG. Of course since the OFI WG’s work is very much in its early stages, there will be quite a bit of discussion devoted to understanding the broad classes of applications that are driving the definition of this next generation of APIs. These classes of applications include data storage applications (both block storage and file or object storage) and applications devoted to data analysis (including Big Data applications such as map-reduce) and even as far as traditional sockets-type applications, in addition to the classical distributed computing environments (such as MPI) that have historically been the forte of the OFA. These discussions should result in a clear consensus on what it means to provide a network stack to support this diverse range of applications.
insideHPC: Is most of you work focused on classical HPC?
OFA: On the contrary! While we are working hard to ensure that whatever interconnect technologies emerge from our efforts address issues in classical HPC, we are also addressing a wider application space. To be specific, we have subdivided this space into the four broad classes of applications we mentioned above: data analytics, which is typically an enterprise application; data access and data storage, two topics of interest to classical HPC but also of intense interest to the enterprise community; distributed computing environments such as MPI or shared memory environments such as PGAS (which tends to be the purview of HPC) and finally classical IP-style applications such as sockets apps which are dearly important in the enterprise. This reflects the OFA’s belief that interconnect technology will continue to play a key role not only in HPC, but also in the broad enterprise space as well.
insideHPC: Where can people find more information on the agenda and how to register?
OFA: Everyone is invited to check out our web site for information on the location and registration. A draft of the agenda is posted there. It will be updated over the coming weeks so people should check back often. The workshop is open to OFA members and non-members.