As Vice President of Software program Engineering, Doug is chargeable for all facets of the Cornelis Networks’ software program stack, together with the Omni-Path Structure drivers, messaging software program, and embedded system management techniques. Earlier than becoming a member of Cornelis Networks, Doug led software program engineering groups at Pink Hat in cloud storage and information companies. Doug’s profession in HPC and cloud computing started at Ames Nationwide Laboratory’s Scalable Computing Laboratory. Following a number of roles in college analysis computing, Doug joined the US Division of Power’s Oak Ridge Nationwide Laboratory in 2009, the place he developed and built-in new applied sciences on the world-class Oak Ridge Management Computing Facility.
Cornelis Networks is a know-how chief delivering purpose-built high-performance materials for Excessive Efficiency Computing (HPC), Excessive Efficiency Information Analytics (HPDA), and Synthetic Intelligence (AI) to main business, scientific, tutorial, and authorities organizations.
What initially attracted you to pc science?
I simply appeared to take pleasure in working with know-how. I loved working with the computer systems rising up; we had a modem at our college that permit me check out the Web and I discovered it fascinating. As a freshman in faculty, I met a USDOE computational scientist whereas volunteering for the Nationwide Science Bowl. He invited me to tour his HPC lab and I used to be hooked. I have been a supercomputer geek ever since.
You labored at Pink Hat from 2015 to 2019, what had been among the initiatives you labored on and your key takeaways from this expertise?
My predominant undertaking at Pink Hat was Ceph distributed storage. I would beforehand targeted totally on HPC and this gave me a possibility to work on applied sciences that had been essential to cloud infrastructure. It rhymes. Lots of the ideas of scalability, manageability, and reliability are extraordinarily related regardless that they’re aimed toward fixing barely completely different issues. By way of know-how, my most necessary takeaway was that cloud and HPC have quite a bit to study from each other. We’re more and more constructing completely different initiatives with the identical Lego set. It is actually helped me perceive how the enabling applied sciences, together with materials, can come to bear on HPC, cloud, and AI functions alike. It is also the place I actually got here to grasp the worth of Open Supply and how one can execute the Open Supply, upstream-first software program growth philosophy that I introduced with me to Cornelis Networks. Personally, Pink Hat was the place I actually grew and matured as a pacesetter.
You’re at the moment the Vice President of Software program Engineering at Cornelis Networks, what are a few of your tasks and what does your common day appear like?
As Vice President of Software program Engineering, I’m chargeable for all facets of the Cornelis Networks’ software program stack, together with the Omni-Path Structure drivers, messaging software program, cloth administration, and embedded system management techniques. Cornelis Networks is an thrilling place to be, particularly on this second and this market. Due to that, I am undecided I’ve an “common” day. Some days I am working with my staff to unravel the most recent know-how problem. Different days I am interacting with our {hardware} architects to ensure our next-generation merchandise will ship for our clients. I am usually within the discipline assembly with our superb group of shoppers and collaborators ensuring we perceive and anticipate their wants.
Cornelis Networks provides subsequent technology networking for Excessive Efficiency Computing and AI functions, might you share some particulars on the {hardware} that’s provided?
Our {hardware} consists of a high-performance switched cloth sort community cloth resolution. To that finish, we offer all the mandatory units to totally combine HPC, cloud, and AI materials. The Omni-Path Host-Material Interface (HFI) is a low-profile PCIe card for endpoint units. We additionally produce a 48-port 1U “top-of-rack” change. For bigger deployments, we make two fully-integrated “director-class” switches; one which packs 288 ports in 7U and an 1152-port, 20U system.
Are you able to talk about the software program that manages this infrastructure and the way it’s designed to lower latency?
First, our embedded administration platform supplies simple set up and configuration in addition to entry to all kinds of efficiency and configuration metrics produced by our change ASICs.
Our driver software program is developed as a part of the Linux kernel. In reality, we submit all our software program patches to the Linux kernel group immediately. That ensures that every one of our clients take pleasure in most compatibility throughout Linux distributions and straightforward integration with different software program akin to Lustre. Whereas not within the latency path, having an in-tree driver dramatically reduces set up complexity.
The Omni-Path cloth supervisor (FM) configures and routes an Omni-Path cloth. By optimizing site visitors routes and recovering shortly from faults, the FM supplies industry-leading efficiency and reliability on materials from tens to 1000’s of nodes.
Omni-Path Categorical (OPX) is our high-performance messaging software program, not too long ago launched in November 2022. It was particularly designed to cut back latency in comparison with our earlier messaging software program. We ran cycle-accurate simulations of our ship and obtain code paths to be able to reduce instruction rely and cache utilization. This produced dramatic outcomes: if you’re within the microsecond regime, each cycle counts!
We additionally built-in with the OpenFabrics Interfaces (OFI), an open commonplace produced by the OpenFabrics Alliance. OFI’s modular structure helps reduce latency by permitting higher-level software program, akin to MPI, to leverage cloth options with out further operate calls.
All the community can also be designed to extend scalability, might you share some particulars on the way it is ready to scale so properly?
Scalability is on the core of Omni-Path’s design ideas. On the lowest ranges, we use Cray link-layer know-how to right hyperlink errors with no latency affect. This impacts materials in any respect scales however is especially necessary for large-scale materials, which naturally expertise extra hyperlink errors. Our cloth supervisor is concentrated each on programming optimum routing tables and on doing so in a fast method. This ensures that routing for even the most important materials might be accomplished in a minimal period of time.
Scalability can also be a essential element of OPX. Minimizing cache utilization improves scalability on particular person nodes with massive core counts. Minimizing latency additionally improves scalability by bettering time to completion for collective algorithms. Utilizing our host-fabric interface sources extra effectively allows every core to speak with extra distant friends. The strategic selection of libfabric permits us to leverage software program options like scalable endpoints utilizing commonplace interfaces.
May you share some particulars on how AI is integrated into among the workflow at Cornelis Networks?
We’re not fairly prepared to speak externally about our inner makes use of of and plans for AI. That mentioned, we do eat our personal pet food, so we get to benefit from the latency and scalability enhancements we have made to Omni-Path to help AI workloads. It makes us all of the extra excited to share these advantages with our clients and companions. Now we have definitely noticed that, like in conventional HPC, scaling out infrastructure is the one path ahead, however the problem is that community efficiency is well stifled by Ethernet and different conventional networks.
What are some modifications that you simply foresee within the {industry} with the appearance of generative AI?
First off, using generative AI will make folks extra productive – no know-how in historical past has made human beings out of date. Each know-how evolution and revolution we’ve had from the cotton gin to the automated loom to the phone, web and past have made sure jobs extra environment friendly, however we haven’t labored humanity out of existence.
By way of the appliance of generative AI, I imagine corporations will technologically advance at a sooner fee as a result of these working the corporate could have extra free time to deal with these developments. For example, if generative AI supplies extra correct forecasting, reporting, planning, and many others. – corporations can deal with innovation of their discipline of experience
I particularly really feel that AI will make every of us a multidisciplinary knowledgeable. For instance, as a scalable software program knowledgeable, I perceive the connections between HPC, huge information, cloud, and AI functions that drive them towards options like Omni-Path. Outfitted with a generative AI assistant, I can delve deeper into the which means of the functions utilized by our clients. I’ve little question that this can assist us design much more efficient {hardware} and software program for the markets and clients we serve.
I additionally foresee an total enchancment in software program high quality. AI can successfully operate as “one other set of eyes” to statically analyze code and develop insights into bugs and efficiency issues. This will probably be significantly fascinating at massive scales the place efficiency points might be significantly troublesome to identify and costly to breed.
Lastly, I hope and imagine that generative AI will assist our {industry} to coach and onboard extra software program professionals with out earlier expertise in AI and HPC. Our discipline can appear formidable to many and it could take time to study to “assume in parallel.” Essentially, identical to machines made it simpler to fabricate issues, generative AI will make it simpler to think about and cause about ideas.
Is there the rest that you simply wish to share about your work or Cornelis Networks normally?
I would prefer to encourage anybody with the curiosity to pursue a profession in computing, particularly in HPC and AI. On this discipline, we’re geared up with probably the most highly effective computing sources ever constructed and we convey them to bear in opposition to humanity’s best challenges. It is an thrilling place to be, and I’ve loved it each step of the way in which. Generative AI brings our discipline to even newer heights because the demand for growing functionality will increase drastically. I am unable to wait to see the place we go subsequent.
Thanks for the good interview, readers who want to study extra ought to go to Cornelis Networks.
