NERSCPowering Scientific Discovery Since 1974

Cori Haswell Nodes

The Cori Phase 1 (also known as the "Cori Data Partition") system is based on Intel Haswell processors and provides approximately the same sustained application performance as Hopper (which was retired in Dec 2015). Cori Phase 1 has an almost identical programming environment to Edison with support for the Intel, Cray and GNU programming environments with the same Cray and Intel high performance math and scientific libraries. The Cori Phase 1 interconnect has the same Aries dragonfly interconnect as Edison.  Cori Phase 1 is different from Edison in that it has twice as much memory per node and has several features to support data-intensive workloads.  Cori Phase 1 is the first NERSC system installed in the new Computational Research and Theory building on the main Berkeley Lab campus, NERSC's new home.  Detailed system configuration information can be found here

System Architecture

  • 1,630 compute nodes
  • 128 GB of memory per node
  • Two 2.3 GHz 16-core HaswellTM processors per node
  • Each core has its own L1 and L2 caches, with 64 KB (32 KB instruction cache, 32 KB data) and 256 KB, respectively; there is also a 40-MB shared L3 cache per socket.
  • Cray Aries high speed "dragonfly" topology interconnect, cabinets, and cooling (same as in Edison)
  • Lustre file system with 30 PB of disk and > 700 GB/second I/O bandwidth
  • Flash 'Burst Buffer' to accelerate I/O performance, a layer of NVRAM that sits between memory and disk.  The Phase 1 system provides approximately 750 GB/second of I/O performance and about 750TB of storage. 
  • > 12 login/interactive nodes
  • SLURM workload manager

Better support for data-intensive science

The Cori Data Partition is designed to accelerate data-intensive applications. Some highlights of this design are:

  • Cray's DataWarp (BurstBuffer) provides high bandwidth and low latency I/O
  • Large amount of memory per node (128 GB/node)
  • Large number of login/interactive nodes to support applications with advanced workflows
  • Immediate access queues for jobs requiring real-time data ingestion or analysis
  • High throughput and serial queues can handle a large number of jobs for screening, UQ, etc.
  • Outbound Internet connection allows your job to communicate with the outside world. For example, you can use this feature to access a database in another center. (Note this is not intended for transferring large data sets.)
  • Two ongoing R&D efforts in collaboration with Cray to maximize Cori's data potential:
    • Enable higher bandwidth transfers in and out of the compute node
    • Enabling docker-like virtualization functionality on cray compute nodes to allow custom software stack deployment