Salishan 2023 Theme
Is HPC at a Crossroads?
Since the advent of Moore’s law in 1965, the high performance computing community has been singularly focused on increasing available computational capabilities with each architectural generation. Arguably, the use of HPL as the standard ruler by which all machines are measured has fueled much of this focus on tera-, peta- and now exa-scale. We’ve spent more than a decade striving for, and now achieving, the goal of fielding an exaflop-capable machine, with more on the near-term horizon. But, for a variety of both technical and budgetary reasons, we haven’t yet been talking about the zetta-scale. Are we in a quandary of what might come next? At this crossroads that we face, is there a clear path on where our community should focus our collective efforts?
In very recent years, the awareness has increased of the opportunity space that the cloud providers and their technologies provide. HPC application developers are dipping their proverbial toes in the water as they explore the pros and cons of offloading some or all of their computational work into the cloud. HPC facilities are increasingly considering the advantages provided by some of the software technologies that cloud providers deploy. And the cloud has embraced hardware disaggregation faster than many of our monolithic procurements.
We’re excited to welcome you back to our first in-person Salishan conference in four years. We plan to have a robust discussion focused on the opportunities and challenges that the HPC community faces if we embrace cloud technologies, including:
In recent years, applications have moved more and more away from traditional, bare-metal, computing to more encapsulated execution models such as computing in the cloud or execution in containers, which separates the application more and more from the actual hardware. In this session, we will explore what impact the move in this direction has for applications and underlying software stack. We will try to weigh user productivity vs. performance vs. reproducibility. And answer questions such as: What compute sizes are more useful to be done in the cloud vs. local? What do we lose by using containers? How much do containers and cloud execution help with reproducibility? Using containers, is it still worth investing time in a full base-metal software stack?
Session 2: Disaggregation and HPC: Promise or Myth
Data Centers are increasingly exploring disaggregation as a solution to deal with growing resource consumption. The promise is higher resource utilization, reduction of stranded resources, improved ability to manage increasing heterogeneity within the data center, better capital and operational cost, and a more adaptable environment in which resources can be composed, i.e., scaled up and out where needed based upon workload requirements without over-provisioning. However, disaggregation requires high levels of intra-resource communication, including stringent requirements for ultra-low latency and ultra-high transmission bandwidth. This state of the technology session poses and will explore the following questions. When, where, and to what extent does disaggregation make sense for HPC systems? Will CXL, a cache-coherent interconnect for data centers, be deployed widely in HPC? Will large-scale supercomputers be disaggregated beyond rack-scale? Should we disaggregate main memory? What are the implications? What is the state of optical I/O?
Session 3: Is the Cloud for Everyone?
Once a “tech-centric” concept, cloud computing is now mainstream within the business world and promises reliable compute capabilities that are readily available and accessible without compromising either performance or security. But for traditional HPC consumers, can the cloud market live up to these promises when compared with traditional HPC offerings? In this session, we will reflect on both the history and current state of the HPC landscape, and discuss the variety of predictions for the future of HPC. We will examine the architectural implications of cloud offerings, and discuss application characteristics that may make cloud computing attractive. Finally, we will look at the use of existing cloud offerings to determine whether cloud computing provides an on-ramp opportunity for new communities to embrace HPC. We seek to answer the questions: Is cloud computing complementary to or a threat to HPC? And Is the cloud for everyone?
Session 4: Opportunities and Challenges in Adopting Cloud Software Technologies for HPC
The technologies and characteristics that uniquely identify high-performance computing (HPC) systems are rapidly diminishing. There are very few hardware technologies that are specific to HPC machines, especially given that the US exascale systems will all be using a high-performance network fabric based on ethernet. The situation with the software infrastructure that enables HPC systems seems to be following the same path. As hyperscalar cloud vendors continue to deploy systems larger than the fastest HPC systems to enable a wide range of resource-constrained workloads, there is a strong desire to leverage as much of this software infrastructure as possible for HPC systems. It's clear that not all of these technologies will work well for traditional HPC applications, but it's also clear that cloud infrastructure can also bring capabilities that HPC systems have long desired, such as increased usability and lower barrier to entry. This session will examine these tradeoffs in depth and provide a perspective on the opportunities and challenges in adopting cloud software infrastructure technologies.
Session 5: Does HPC / Computing Have a Role in Edge?
Edge computing, wherein compute resources are placed in proximity to geographically distributed data sources, is an increasingly common paradigm across science and national security mission applications. This session will explore opportunities and challenges at the intersection of HPC and the edge: what science and mission applications benefit from the intersection of HPC and the Edge? What policy, cultural, and technical challenges need to be overcome before these applications can be realized? Given industry investment at the intersection of AI, edge, and cloud computing, what lessons and tools can be brought to bear for the HPC community?
Since the advent of Moore’s law in 1965, the high performance computing community has been singularly focused on increasing available computational capabilities with each architectural generation. Arguably, the use of HPL as the standard ruler by which all machines are measured has fueled much of this focus on tera-, peta- and now exa-scale. We’ve spent more than a decade striving for, and now achieving, the goal of fielding an exaflop-capable machine, with more on the near-term horizon. But, for a variety of both technical and budgetary reasons, we haven’t yet been talking about the zetta-scale. Are we in a quandary of what might come next? At this crossroads that we face, is there a clear path on where our community should focus our collective efforts?
In very recent years, the awareness has increased of the opportunity space that the cloud providers and their technologies provide. HPC application developers are dipping their proverbial toes in the water as they explore the pros and cons of offloading some or all of their computational work into the cloud. HPC facilities are increasingly considering the advantages provided by some of the software technologies that cloud providers deploy. And the cloud has embraced hardware disaggregation faster than many of our monolithic procurements.
We’re excited to welcome you back to our first in-person Salishan conference in four years. We plan to have a robust discussion focused on the opportunities and challenges that the HPC community faces if we embrace cloud technologies, including:
- HPC applications in the cloud: Gains and losses
- Disaggregation and HPC: Promise or Myth
- Is the Cloud for Everyone?
- Opportunities and Challenges in Adopting Cloud Software Technologies for HPC
- Does HPC / computing have a role in edge?
In recent years, applications have moved more and more away from traditional, bare-metal, computing to more encapsulated execution models such as computing in the cloud or execution in containers, which separates the application more and more from the actual hardware. In this session, we will explore what impact the move in this direction has for applications and underlying software stack. We will try to weigh user productivity vs. performance vs. reproducibility. And answer questions such as: What compute sizes are more useful to be done in the cloud vs. local? What do we lose by using containers? How much do containers and cloud execution help with reproducibility? Using containers, is it still worth investing time in a full base-metal software stack?
Session 2: Disaggregation and HPC: Promise or Myth
Data Centers are increasingly exploring disaggregation as a solution to deal with growing resource consumption. The promise is higher resource utilization, reduction of stranded resources, improved ability to manage increasing heterogeneity within the data center, better capital and operational cost, and a more adaptable environment in which resources can be composed, i.e., scaled up and out where needed based upon workload requirements without over-provisioning. However, disaggregation requires high levels of intra-resource communication, including stringent requirements for ultra-low latency and ultra-high transmission bandwidth. This state of the technology session poses and will explore the following questions. When, where, and to what extent does disaggregation make sense for HPC systems? Will CXL, a cache-coherent interconnect for data centers, be deployed widely in HPC? Will large-scale supercomputers be disaggregated beyond rack-scale? Should we disaggregate main memory? What are the implications? What is the state of optical I/O?
Session 3: Is the Cloud for Everyone?
Once a “tech-centric” concept, cloud computing is now mainstream within the business world and promises reliable compute capabilities that are readily available and accessible without compromising either performance or security. But for traditional HPC consumers, can the cloud market live up to these promises when compared with traditional HPC offerings? In this session, we will reflect on both the history and current state of the HPC landscape, and discuss the variety of predictions for the future of HPC. We will examine the architectural implications of cloud offerings, and discuss application characteristics that may make cloud computing attractive. Finally, we will look at the use of existing cloud offerings to determine whether cloud computing provides an on-ramp opportunity for new communities to embrace HPC. We seek to answer the questions: Is cloud computing complementary to or a threat to HPC? And Is the cloud for everyone?
Session 4: Opportunities and Challenges in Adopting Cloud Software Technologies for HPC
The technologies and characteristics that uniquely identify high-performance computing (HPC) systems are rapidly diminishing. There are very few hardware technologies that are specific to HPC machines, especially given that the US exascale systems will all be using a high-performance network fabric based on ethernet. The situation with the software infrastructure that enables HPC systems seems to be following the same path. As hyperscalar cloud vendors continue to deploy systems larger than the fastest HPC systems to enable a wide range of resource-constrained workloads, there is a strong desire to leverage as much of this software infrastructure as possible for HPC systems. It's clear that not all of these technologies will work well for traditional HPC applications, but it's also clear that cloud infrastructure can also bring capabilities that HPC systems have long desired, such as increased usability and lower barrier to entry. This session will examine these tradeoffs in depth and provide a perspective on the opportunities and challenges in adopting cloud software infrastructure technologies.
Session 5: Does HPC / Computing Have a Role in Edge?
Edge computing, wherein compute resources are placed in proximity to geographically distributed data sources, is an increasingly common paradigm across science and national security mission applications. This session will explore opportunities and challenges at the intersection of HPC and the edge: what science and mission applications benefit from the intersection of HPC and the Edge? What policy, cultural, and technical challenges need to be overcome before these applications can be realized? Given industry investment at the intersection of AI, edge, and cloud computing, what lessons and tools can be brought to bear for the HPC community?