National Supercomputing CenterNURION
5th Supercomputer Summary
Nurion is a system consisting of compute nodes, CPU-only nodes, Omni-Path interconnect networks, Burst Buffer high-speed storage, Luster-based parallel file system, and water cooling device based on Rear Door Heat Exchanger (RDHx). Nurion's compute nodes are 8,305 Intel Xeon Phi processors (named "Knight Landing") nodes and CPU-only nodes are 132 Intel Xeon processors (named "Skylake") nodes. Total theoretical performance is 25.7 petaflops, which was ranked 11th in the world in June 2018 (http://www.top500.org).
| Compute Node | CPU-only Node | |
|---|---|---|
| Model | Cray CS500 | |
| Architecture | Xeon Phi cluster | CPU cluster |
| Processor | Intel Xeon Phi 7250 1.4GHz | Intel Xeon 6148 2.4GHz |
| Total nodes | 8,305 | 132 |
| CPU each node | 1 | 2 |
| Cores per CPU | 68 | 20 |
| Cores each node | 564,740 | 5,280 |
| Memory per core | 1.4GB | 2.4GB |
| Memory each node | 96GB | 192GB |
| Total memory | 797.3TB | 25.3TB |
| Compute power(peak) | 25.3PF | 0.4PF |
| Storage capacity | 21PB | |
| Interconnect | OPA | |
Nurion features a theoretical performance of 25.7PFlops, 70 times better capacity than 4th supercomputer and manycore processor with high performance interconnects and storage to provide high quality service environments for data-intensive applications by introducing burst buffer storage as well as compute-intensive applications.
Burst Buffer
- It acts as an I/O buffer by storing bursty I/O request traffic, which occurs instantly in the application program, between the application and the parallel file system through high-performance hardware and software
- Recently, it is becoming more popular for large supercomputers to adopt burst buffers for high performance I/O layers using NVMe SSD.
5th Supercomputer Performance
| Category | Features | # of nodes | Score | World Ranking | |
|---|---|---|---|---|---|
| HPL | Large-scale Dense Matrix Computation Used for Top500 |
8,174(KNL) + 122(CPU-only) | 13.93PF | 11th (Jun 2018) |
|
| HPCG | Large-scale Sparse Matrix Computation Similar to normal user applications |
8,250(KNL) | 0.39PF | 8th (Nov 2018) |
|
| IO500 | Various IO Workloads | 2,048(KNL) | 160.67 | 2nd (Nov 2018) |
|
| GRAPH500 | BFS | Breadth-First Search, Interconnect Performance | 1,024(KNL) | 1456.46GTEPS | 9th (Nov 2019) |
| SSSP | Single Source Shortest Path | 1,024(KNL) | 337.239GTEPS | 3rd (Nov 2019) |
|
Certificate
Nurion system configuration and figures
-
Main system
- 8,305 compute nodes
- 132 CPU-only nodes
(25.7PFlops total)
-
Storage
- 12 racks(23.88PB total)
- 3,200 8TB NL-SAS
- 768 1.2TB SSDs
-
Tape Library
- 4 racks(10PB total)
- 1,700 LTO7 tapes
-
Infrastructure
- Chiller 400RT 4EA
- Free colling 250RT 2EA
- Cooling tower 450RT 4EA
- HVAC 30RT 11EA
