Back to Warp Speed, Scotty!

Aren’t supercomputers already supposed to be moving at “warp” speed? Well, multi-petaflops anyway?

With computing challenges getting more demanding and datasets getting more massive, it is not as easy as Captain James Kirk asking Scotty for more power. Supercomputers and their associated storage systems are getting larger each year, but a performance inefficiency is emerging between the compute nodes and the disk systems where the parallel file systems are located.

The overall performance of application execution on supercomputing systems is subject to the efficiency of the data movement between a tier of spinning disks and the memory where the codes are actually executing.

Cray has been collaborating with some of the most demanding HPC users in the industry to better understand the new choke points that emerge when supercomputers and disk-based storage systems grow at a faster rate than the evolution of the intercommunications between them. A new set of innovations is required to address this inefficiency gap in supercomputing storage systems to enable muliti-petascale operation today and continued progression toward exascale tomorrow.

The introduction of the new Cray® XC40TM supercomputer delivers on the first element of that challenge – it doubles the scalable performance of previous-generation Cray XC series supercomputers. Integrating the new Intel Xeon processor E5-2600 v3 (previously code named Haswell) into the HPC-optimized computing blades with faster DDR4 memory per compute node, the XC40 supercomputer is two times faster than the XC30 systems, while leveraging the same proven Cray software stack, Aries interconnect and Dragonfly network.

Thanks to an “early ship” program, Cray XC40 supercomputers are already installed at numerous customer sites around the world, and Cray has announced many other high-profile contracts for systems that were publicized as “next-generation XC series systems” prior to the XC40 product announcement.

The other innovative element addresses total system storage cost reduction as well as the performance efficiencies gap between the compute and disk tiers. Many datacenters are overprovisioning their storage systems to ensure they can meet the peak or worst-case data I/O activity of their demanding compute jobs, spending several times more on spinning disks than they would use in average operation. These extra, seldom-used resources have the obvious capital equipment expenses, but also can hide excessive operating costs of power consumption and floor space footprint.

Here is where Scotty pushes the thruster control toward warp speed.

Cray has introduced a new tier of fast flash storage that can absorb the spikes of peak I/O demand for a diverse variety of use cases, and eliminate the emerging performance gap. The Cray DataWarpTM applications I/O accelerator provides node-local SSDs of flash storage and management software (smart I/O forwarding, virtualization, workload manager integration and more). Unlike other industry implementations of SSD integration in the disk system, all XC series compute and I/O blades benefit from the fast and flexible Aries interconnect, and the Cray fast flash-powered I/O blades are inserted directly into the XC series supercomputer banks of compute blades to move that performance as close to the application execution as possible.

Applications with heavy I/O activity will benefit immediately from this new fast cache layer for streamlining the interactions between the parallel file system (PFS) and the compute nodes. And larger systems will see tremendous acceleration of checkpoint/restart operations due to the decoupling of the fast DataWarp snapshots and asynchronous transferring of that data to the PFS, enabling the application codes to continue executing with reduced communications overhead.

Using the customer’s unique requirements and balancing the system configuration across three tiers of compute memory, DataWarp flash, and PFS on spinning disks, we can improve system performance while reducing both total capital costs and operating costs.

Scotty … we do have the power!

Speak Your Mind

Your email address will not be published. Required fields are marked *