HPE Slingshot interconnect

A high rate of small packets is an increasingly important metric for many applications. Supporting link speeds of 200 Gb/s, HPE Slingshot provides state-of-the-art small packet performance on standard links with more than 1.2 billion packets per second per port (600M packets per second each direction). The HPE Slingshot interconnect network delivers high packet rates on IP, messaging, and remote memory access.

The ability to manage network resources is highly developed in scale-out data centers but absent from many HPC systems. HPE Slingshot provides system-wide quality of service (QoS) classes. With these comes the ability to control how network bandwidth is allocated. Individual jobs can be assigned to different QoS classes ensuring high priority jobs get the network resources they require. Multiple classes can even be assigned to different traffic classes (bulk vs. low latency as example) within a job.

Rather than routing traffic along a predetermined path, HPE Slingshot intelligently routes packets dynamically according to load. The route is determined using real-time, global information on load and is gathered by the hardware and distributed among the switches providing an up-to-date picture of which paths to use or avoid. This enhanced approach provides both adaptive routing of ordered flows, and packet-by-packet adaptive routing of traffic that does not require ordering.

HPE Slingshot employs a revolutionary congestion management mechanism that provides strong performance isolation between workloads, limiting the impact of poorly behaved applications on production traffic and system services. Congestion management is performed automatically in hardware, quickly identifying sources of congestion and limiting their ingress, while allowing other traffic to continue flowing. The mechanism is stable, quick to converge, robust across a wide variety of traffic patterns, and suitable for the highly dynamic communication patterns found in today’s HPC and converged workloads.

To enhance interoperability with storage architectures and datacenters, HPE Slingshot was designed to be Ethernet compatible. This means HPE Slingshot switches can connect directly to third-party Ethernet-based storage devices as well as to datacenter Ethernet networks. Applications running on HPE Cray supercomputers and HPE HPC 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’s high radix 64 port switch coupled with the Dragonfly topology scales can scale to over 250,000 endpoints with a maximum of three switch-to-switch hops between any endpoints. In addition to latency improvements in larger installations, this low-diameter network reduces network equipment, cabling, and power and cooling costs. It also facilitates the use of innovative adaptive routing algorithms that improve application performance.