Cray revealed today the details of its next-gen supercomputing architecture, Shasta, selected to be the next flagship system at NERSC. We’ve known of the code-name “Shasta” since the Argonne slice of the CORAL project was announced in 2015 and although the details of that plan have changed considerably, Cray didn’t slow down its timeline for Shasta. Set for commercial launch in late 2019, the exascale-class architecture unifies the Cray supercomputer and cluster product lines and debuts a brand-new Cray-designed system interconnect, called Slingshot.

In tandem with Cray’s Shasta disclosures, the U.S. Department of Energy (DOE) is announcing that NERSC, the National Energy Research Scientific Computing Center, has chosen a Cray Shasta supercomputer for its NERSC-9 system, slated for delivery in late 2020. Named “Perlmutter” (after Nobel Prize winning astrophysicist Saul Perlmutter), the system will feature AMD Epyc processors and Nvidia GPUs offering a combined peak performance of ~100 petaflops and a sustained application performance equivalent to about 3X that of the Cray Cori (NERSC-8) supercomputer. The new contract, which includes a Cray Clusterstor storage system, is worth a reported $146 million, one of the largest in Cray’s history. (We’ll be reporting additional details of the NERSC-9 system soon.)

Shasta will be the first Cray architecture to support multiple cabinet types, a 19” air- or liquid-cooled, standard datacenter rack and a high-density, liquid-cooled rack designed to hold 64 compute blades with multiple processors per blade. Both options can scale to well over 100 cabinets, according to Cray.

Currently Cray’s product set is divided between its CS line — dating to its Appro acquisition and featuring commodity software, a third-party interconnect and a standard 19” rack — and its flagship XC line, the dense, scale-optimized, liquid-cooled rack offering that integrates its custom Aries interconnect and a custom software stack. Shasta breaks this convention and allows the user to have the Cray custom interconnect and software stack in either form factor. Shasta also emphasizes infrastructure choice, offering from one to 16 nodes per compute blade, support for x86 (Intel and AMD), Arm (Marvell), GPUs (Nvidia and AMD) in the same system, as well as system interconnects from Cray (Slingshot), Intel (Omni-Path) or Mellanox (InfiniBand). Cray also supports FPGAs and anticipates offering support for one or more of the emerging ML accelerators in the next one-to-two years.

“We designed Shasta with both extreme performance and flexibility,” comments Cray CTO Steve Scott in an interview with HPCwire. “The new system architecture is motivated by increasingly heterogeneous data-centric workloads. We’re seeing more and more customers wanting to run workflows containing simulation and analytics and AI, and we needed systems that could handle all of these simultaneously, moving away from the siloed situation of different systems optimized for different workloads.”

The new architecture has the ability to handle high processor power levels with direct liquid cooling, supporting W4-class warm water cooling (up to 45-degree Celsius water). Per cabinet cooling tops out at 250 kilowatts initially, increasing to 300 kilowatts per cabinet within the first year after launch, according to Cray.

Introducing Slingshot, Cray’s 8th Generation HPC Network

One of the biggest hardware revelations in the Shasta unveiling is a new interconnect, called Slingshot, designed to scale to exascale and beyond with support for over 250,000 end points. The heart of slingshot is a 64-port switch with 200 Gbps ports (based on 50 Gbps signalling technology), providing 12.8 Tbps bandwidth per switch. Slingshot implements the Dragonfly topology, which Cray invented in 2008. Cray reduced the network diameter from five hops in the current Cray XC generation Dragonfly topology, to three hops for Slingshot, with a reported latency of around 300 nanoseconds per hop.

“Fewer hops results in lower latency, reliability goes up, and the adaptive routing gets that much better because at the place where you’re making the routing decisions you’ve got a pretty good idea of the global state of the network because of the really low diameter and because our switches exchange information with each other about the state of congestion in the network,” says Scott.

Slingshot is said to be highly configurable in that it can accommodate different node sizes and can vary the amount of network injection bandwidth to match workloads. Cray also made Slingshot Ethernet compatible to enable interoperability with third-party storage devices and datacenters in support of today’s greater need for data exchange between platforms.

Scott is especially proud of Slingshot’s novel congestion control mechanism, claimed to dramatically reduce queuing latency in the network. Scott says this is where, in practice, most latency actually comes from and the mechanism provides strong performance isolation between workloads. He adds that this is something that’s notoriously difficult to do and no one has done it yet for HPC workloads.

“There are existing congestion control mechanisms designed for datacenters and they are particularly difficult to tune. They are fragile. They are slow to converge and they just don’t work well for HPC workloads. We’ve cracked the code and figured out how to do this with Slingshot,” says Scott. “It means you’re going to get low latency, both for average latency and tail latency (the latency that the slowest 1 percent or tenth of a percent of packets experience). In traditional HPC systems, Cray’s current systems included, one workload that causes congestion can really interfere with other workloads running on the system and cause those latencies and especially those tail latencies to go off, so Slingshot’s really going to provide performance isolation and provide low and consistent network latency.”

All Together Now

Scott adds that another key part of their data-centric design is having a really strong I/O and storage system. With Shasta, Cray takes the storage system, which typically has been external to the supercomputer, and pulls it into the supercomputer directly onto the high-speed Slingshot network, obviating the need for L net router nodes and external InfiniBand network. “We will have high-performance flash and hard drive based storage enclosures that attach directly onto the Slingshot network. This reduces complexity, reduces latency and really improves the performance, especially for fine-grained I/O,” says Scott, adding that a high-performance flash tier plus a high capacity hard-drive based tier will be unified in the same Lustre namespace with tiering between them.

“Cray is widely seen as one of only a few HPC vendors worldwide that is capable of aggressive technology innovation at the system architecture level,” said Steve Conway, Hyperion Research senior vice president of research, in support of today’s announcement. “Cray’s Shasta architecture closely matches the wish list that leading HPC users have for the exascale era, but didn’t expect to be available this soon. This is truly a breakthrough achievement.”

Another vote of confidence came from Dr. Sudip Dosanjh, director of the NERSC Center at Lawrence Berkeley National Laboratory, where a Shasta system with a mix of AMD Epyc CPUs and Nvidia GPUs will be supporting a diverse set of HPC workloads.

“Our scientists gather massive amounts of data from scientific instruments like telescopes and detectors that our supercomputers analyze every day,” said Dosanjh. “The Shasta system’s ease of use and adaptability to modern workflows and applications will allow us to broaden access to supercomputing and enable a whole new pool of users. The ability to bring this data into the supercomputer will allow us to quickly and efficiently scale and reduce overall time to discovery.  We value being able to work closely with Cray to provide our feedback on this next generation system which is so critical to extending our Center’s innovation.”

It’s been a few years and a few technology generations since we’ve seen AMD+Nvidia together in a leadership-class system, but Cray notes that there was no technical challenge or special work required to use AMD CPUs with Nvidia GPUs. “Both are designed well to work with other components, and Cray’s programming environment is well suited for targeting them,” a company spokesperson told us.

Cray expects to share specific product information and system names next spring in anticipation of making Shasta systems commercially available by the end of 2019. Cray will be showcasing Shasta and Slingshot next month (Nov. 11-16) at the 30^th anniversary of the Supercomputing (SC) conference in Dallas, Texas.