Arctur-2: The reincarnation of Dawning 3000

A long time ago — in the year of 2001, in a land far far away — the People’s Republic of China; a new HPC system was born — the Dawning 3000. Today — in the year of 2016, in the centre of Europe, a new star was born: the Arctur-2 hyper-converged HPC system — designed and developed by Arctur d.o.o..

Dawning 3000, 2001

You’ll probably ask: what does these two system have in common? Well, surprisingly enough, the hardware for both systems were made by the same manufacturer — Dawning Information Industry which is now known as Sugon Information Industry. And even though 15 years has passed since Dawning 3000, we, the HPC engineers, are still seeking to reach the same goals: to make an efficient and fast high-performance computing system.

Arctur-2, 2016

Okay, to be honest, there aren’t really that many similarities between the two systems (yes, 15 years feels like an eternity in HPC). But there is one interesting decision in Arctur-2 rack design that really does remind us of the Dawning 3000 — the circular placement of server racks. This kind of placement was chosen for Arctur-2 mainly to achieve exceptional cooling efficiency. Hot air is captured in the middle of circle, sucked up above the ceiling and exhausted outside — while returning fresh outside air in front of the racks. This enabled us to reach exceptional Power Usage Efficiency (PUE) factor of 1.1 with minimal investment in cooling technology — based on passive cooling with outside air.

As far as performance is concerned — well, this is where the “eternity feeling” comes into play. Dawning 3000 was capable of reaching a total combined speed of 403.2 GFLOPS for the whole system. A single CPU, used in Arctur-2 (Intel E5–2690 v4, 14 core) is capable of 470 GFLOPS — yes, a whole Dawning 3000, smashed into a single CPU! There are 72 CPUs in Arctur-2 with a combination of Intel E5–2690 v4, 14 core and Intel E5–2650L v4, 14 core, giving a total theoretical power of 30 TFLOPS (30.000 GLOPS).

Anyway, the main strength of Arctur-2 is not it’s pure theoretical compute power, but it’s versatility. Arctur-2 is a flagship hyper-converged HPC & Cloud computing infrastructure. Is has been designed and built to fit various and diverse usage scenarios, from general-use Cloud Computing to small-scale HPC computing as well as small-scale Big Data analytics. By converging different types of servers into single entity we achieved high performance, energy efficiency and seamless management, enabling us to provide excellent services to our customers.

Brief technical details reveals the strengths of Arctur-2. A total count of 1008 CPU cores from Intel Xeon E5 v4 CPUs gives enough pure compute power. As much as 1 TB of memory per node and up to 14 locally attached hard drives makes ideal Big Data crunching machine. Combined with virtual ScaleMP software aggregation we’re able to provide SMP system with 8 TB memory and 224 CPU cores — ideal for many industry in-memory or memory-demanding applications. And there are GPU nodes with NVIDIA Tesla GPUs as well — 4 of them per node, each dual-GPU which gives you as much as 8 GPUs per single server — a dream for any CUDA or OpenCL developer. We also decided to go cutting-edge with networking. Forget InfiniBand, here is the latest 100 Gigabit Ethernet network from Mellanox which we deployed cluster-wide. Also, the whole system is backed with 1.5 PB of distributed storage, based on Ceph tehnology. Cloud computing (virtual machines) environment is backed with multiple-tier block storage arrays including all-flash arrays for most demanding workloads.

The mission of Arctur-2 is to serve the industry, to unleash the power of small and medium businesses and to bring powerful computing to the engineers. I bet it will do it’s mission just great.

Marko Kobal, PhD.