HPE Slingshot interconnect redefines performance for HPC clusters

[AdvEXperts]

Previously limited to HPE supercomputers, you can now get the performance and interoperability of the groundbreaking HPE Slingshot interconnect for HPC clusters of all sizes.

At SC21 this November, we announced we would be extending availability of HPE Slingshot—our groundbreaking HPC interconnect solution—from HPE Cray supercomputers to HPC clusters as well.

That moment is here. We’re pleased to announce that HPE Slingshot is available for HPE Apollo and HPE DL ProLiant HPC cluster solutions. Now HPC clusters of all sizes can take advantage of HPE Slingshot performance and interoperability.

Increasingly, our HPC users want to run a mix of workflows (like simulation, analytics, and AI) on one system that can handle them all simultaneously.  This means accessing more types of data from the data center and serving a more diverse range of software that evolved in the cloud, not in the world of HPC. Our HPC users do not want to continue addressing different workloads with different systems, and HPE Slingshot was designed to support a flexible, heterogeneous architecture that handles the increasingly data-centric, converged and AI-focused workloads.

What makes HPE Slingshot so effective

Along with many new performance-focused features including an extremely high bandwidth switch with 64 ports operating at up to 200 Gb/s, HPE Slingshot is Ethernet based. This means that HPE Slingshot switches can connect directly to third-party Ethernet-based storage devices and to data center Ethernet networks.

Applications running on HPE Slingshot-connected clusters can directly exchange IP/Ethernet traffic with the outside world, making it easier and more efficient to ingest data from external sources—an increasingly important consideration in this highly networked and data-driven world.

HPE Slingshot is a revolutionary new interconnect that redefines performance. And not just great performance—no-compromise performance, in real world use, with many diverse applications sharing the system. It combines the best of traditional HPC networking solutions with the best of Ethernet, delivering the benefits of an Ethernet-based network to seamlessly run industry standard Ethernet-based applications, while at the same time delivering the latency and bandwidth HPC workloads require.

Theoretical bandwidth and latency in real-world deployments

The latency and bandwidth that matters in HPC isn’t the theoretical minimal latency, which represents an idle network, or the theoretical bandwidth through the network’s bisectional bottleneck.  It is the real latency and bandwidth achieved under load, and at scale, which can degrade significantly with overloaded interconnect paths due to unbalanced traffic patterns.

HPE Slingshot incorporates unique technology innovations to consistently ensure theoretical bandwidth and latency are maintained in real world deployments. Every switch in the switch fabric understands the configuration of all switches and gets real time information on traffic flow.  Individual packets route through the switch network on the optimal path, based on information including link congestion, available bandwidth, hop count, errors, degradation, or outages. This ensures continuous load balancing across the links and maximizes the realized system bandwidth for both ordered and unordered traffic. 

Advanced congestion management

But even with adaptive routing, diverse traffic patterns and multiple workloads competing for the same resources can mean that traffic to an endpoint can exceed what it can handle—resulting in congestion. Advanced congestion management innovations that are fully automatic and implemented in hardware quickly distinguish between the traffic causing congestion and the traffic affected by it. The network then automatically regulates the flow of congesting traffic, allowing it to make progress with minimal impact to the other traffic.

Key to reducing overall costs

HPE Slingshot is also designed to efficiently deliver this performance with less networking infrastructure which is key to reducing overall costs of an HPC cluster solution.  It’s high radix, 64-port switches reduce switch count and total switch cost.  Supported NICs include third party Ethernet PCIe cards and our own HPE Slingshot 200 Gbit/sec PCIe card.  Our Dragonfly topology uses about half of the optical cable that would be required for other topologies for the same global bandwidth which reduces both cost and power.

Built-in agility

Finally, HPE Slingshot was designed with a robust feature set with agility for the future.  With an HPE Slingshot based cluster, you can be confident you are ready for years for whatever users will run and that the performance of the interconnect can handle it.  IP workloads run natively alongside the highest performing RDMA codes to help you manage the software your HPC cluster users will implement.  Ethernet based, HPE Slingshot connects directly to existing storage and new sources of data in the future, such as instruments and sensors, to let you easily add data into your infrastructure.

The HPE Slingshot interconnect combines the best of traditional HPC networking solutions with the best of Ethernet, delivering consistent, reliable performance across a broad range of both workloads and system sizes. It enables a new class of users to transition from traditional cloud-based computing and take advantage of supercomputing class performance and capabilities and at any scale.

For more information on how you can implement the HPE Slingshot interconnect, contact your HPE representative or visit www.hpe.com/slingshot.

---

Meet HPE blogger Marten Terpstra, Sr. Director of Product Management for High Performance Networking 

Marten Terpstra is a Senior Director of Product Management for High Performance Networking at HPE’s HPC and AI Business Unit. He is focused on developing innovative fabric and system interconnect technologies and solutions for High Performance Computing, Artificial Intelligence and high performance Data Solutions, including HPE Slingshot, the industry’s only Ethernet based HPC and AI fabric.

Advantage EX Experts
Hewlett Packard Enterprise

twitter.com/hpe_hpc
linkedin.com/showcase/hpe-ai/
hpe.com/info/hpc