The Australian Research Council-funded FlashLite was installed in Queensland’s Polaris Data Centre in August last year. FlashLite set a record CFP2006 Rate result of 33,000, and CINT2006 Rate result of 40,200, beating previous top position SGI UV systems by as much as 29 per cent.
FlashLite also reached a STREAM Triad benchmark result of 6,367 GB/s, taking the number three position on the STREAM Top 20 list, trailing only two systems which have double the processor count.
FlashLite, installed by XENON Systems, was configured as an SMP of 128 processors having 1,536 cores and 16 TB RAM, powered by ScaleMP’s vSMP Foundation and running Linux 64bit (CentOS 6.5).
“We are extremely satisfied to have set two world records,” said Prof. David Abramson, Research Computing Centre Director. “These benchmarks measure the computational power and the speed of accessing memory respectively. FlashLite was designed for data-intensive science, but it’s great to see it achieving outstanding computational performance at the same time. This makes it ideal for demanding Big Data problems such as those found in Life Sciences, Physics, Chemistry, Geo-Science and even Social Science.”
Shai Fultheim, ScaleMP’s CEO and founder, said, “With the FlashLite system, the University has proven that one can indeed use commodity components to create the world’s fastest shared memory computer, and we look forward to further collaboration with Queensland scientists in varying fields of research, to help them get the most out of their modeling, simulation, and analysis of the universe we live in.”
“These fantastic world records confirm the capabilities of this machine, and XENON’s integration of vSMP Foundation makes it easy to partition and reconfigure the system as needed by the researchers,” said Dragan Dimitrovici, founder and CEO of XENON Systems.
FlashLite has been designed to support data intensive applications, which are neither well served by traditional supercomputers nor by modern cloud-based data centres. Conventional supercomputers maximise Floating Point Operations per Second (FLOPS) and inter-processor communication rates through high bandwidth and low latency networks. Conversely, modern cloud systems minimise the cost of ownership through reliance on virtual machines and shared storage; they thus utilise relatively slow processors and networks and, by and large, do not support parallel processing.
FlashLite, however, maximises Input Output Operations Per Second (IOPS) while achieving competitive FLOPS ratings and high performance networking, producing a balanced system for applications that exploit parallelism, high-speed arithmetic, and high performance Input/Output (IO). The machine also uses ScaleMP's vSMP software that provide seamless access to data regardless of location, making it easier to build new data intensive applications, whilst supporting legacy codes using familiar techniques.
FlashLite is a multi-node cluster being purchased from XENON Systems that is comprised of the following sub-systems:
- 68 x compute nodes
- 2 x login nodes and 2 x administration nodes
- dual rail 56Gbps Mellanox infiniband fabric
- non-blocking within groups of 24 nodes
- 2:1 blocking factor between groups of 24 nodes
- ScaleMP vSMP software that aggregates multiple nodes into "super nodes" with larger memory/CPU/disk/IO than the individual nodes
- ROCKS cluster management software
- Torque + Maui batch system
- 150+ TB of high performance storage connected via NFS into IB fabric.
- Each compute node has the following attributes:
- 2 x Xeon E5-2680v3 2.5GHz 12core Haswell processors with 30MB Smart Cache
- 16 x 32GB DDR4-2133 ECC LRDIMM modules – total 512GB (256GB per socket)
- 2 x 500GB 2.5" 7.2K HDDs as RAID 1 system disk
- 3 x 1.6TB Intel P3600 2.5" NVMe (SSD) drives for local data storage
- 2 x Mellanox Connect-IB 56Gb/s FDR Single Port Infiniband PCIe3 x8 adapter.
- The login nodes are identical to the compute nodes except that they have:
- 0 x 1.6TB Intel P3600 2.5" NVMe (SSD) drives
- 4 x 480GB Intel S3500 SSD drives for local data storage.
FlashLite has been funded by the Australian Research Council, in conjunction with the following stakeholders:
- Griffith University
- Monash University
- Queensland Cyber Infrastructure Foundation (QCIF)
- Queensland University of Technology
- The University of Queensland
- The University of Technology, Sydney.
SPEC benchmarks were executed by ScaleMP, Inc., between 12–14 January 2016. Comparisons are according to the Standard Performance Evaluation Corporation’s CPU 2006 benchmark results, published by SPEC on 17 February 2016.