The Electronics and Telecommunications Research Institute (ETRI) has developed a technology that shares idle memory across servers in a data center to overcome memory capacity limits on individual servers.
ETRI said on Wednesday it has applied its in-house data center memory hypervisor, EMP (Elastic Memory Platform), to a 500-node high-performance computing (HPC) cluster at the University of Southern California (USC) Advanced Research Computing Center (CARC) and has been running a pilot service since September.
EMP is system software that links memory distributed across multiple servers through an ultra-high-speed network and provides it as expandable memory to servers that need it. Even if a particular server runs short of memory, it can use unused memory on other servers to perform large-scale computations.
In an actual performance verification, it succeeded in running a genome analysis task requiring about 1.8 TB of memory on a server with 448 GB of local memory. The server’s own memory was about 25 percent of the required capacity, but it secured the shortfall by using memory on other servers connected through an InfiniBand ultra-high-speed network.
In a performance test on bio application programs requested by USC CARC, it recorded up to 2.30 times higher performance per unit of memory than the existing system. The application also laid a foundation to use up to 384 TB of memory resources in USC CARC’s HPC cluster through EMP.
EMP can be applied without modifying the source code of existing applications. ETRI has completed integration with SLURM, an HPC standard job scheduler. It is also pushing to integrate with Open OnDemand, a web-based user interface.
ETRI plans to further verify performance and usability with USC researchers in large-scale scientific computing fields such as genome assembly, metagenome data analysis, protein analysis and weather forecasting. It will expand use of the technology to domestic HPC centers and universities and research institutes and pursue commercialisation.
Ko Kwang-won (고광원), head of ETRI’s AIDC Research Division, said EMP is a technology that shares idle memory on existing servers to overcome individual server memory capacity limits and enable large-scale computations. He said ETRI will develop it into core system software that raises resource utilisation efficiency at domestic AI data centers.