[Prof. Myoungsoo Jung, PHD candidate Donghyun Gouk, PHD candidate Miryeong Kwon, From left]
Our department’s Professor Myounsoo Jung’s research team has developed the world’s first CXL2.0 based freely scalable and direct accessible memory expanding platform DirectCXL.
The research team has demonstrated the large-size datacenter applicationon on the end-to-end memory expanding platform consisting CXL hardware prototype and operating system. Though a few of the memory vendors just showed a single memory device, it is the first to demonstrate the application on the full platform with operating system. Compared to conventional memory expanding system, DirectCXL shows 3x performance improvement in executing data center application and supports increasing the memory capacity greatly.
RDMA based memory expanding solution which is commonly used in data center can expand system’s memory by adding memory node which consist of CPU and memory. However, the RDMA solution degrades the performance and needs a substantial budget to add memory node with CPU. To address these problems, PCI express interface based new protocol called CXL which supports high performance and scalability has appeared, but memory vendors and academia fall on hard times in conducting the research into CXL.
To suggest the solution and cornerstone about CXL2.0 based memory expanding, Jung’s research team developed CXL memory device, host CXL processor and CXL network swith to expand system’s memory. They also developed Linux based CXL software module so that existing computer system can control these memory expanding platform. With our proposed DirectCXL, memory capacity can be scaled out freely without extra cost of computing resources.
This work is expected to be utilized in a variety of ways, such as data centers and high-performance computing, as it can provide efficient memory expanding and high performance.
The paper (Direct Access, High-Performance Memory Disaggregation with DirectCXL) was reported in July, 11th at ‘USENIX Annual Technical Conference, ATC, 2022’.
In addition, the research was introduced to the UK top technology newspaper ‘The Next Platform’ with Microsoft and Meta(Facebook)(https://www.nextplatform.com/2022/07/18/kaist-shows-off-directcxl-disaggregated-memory-prototype/) and will be presented in August 2nd/3rd at CXL forum in Flash Memory Summit.
More information about ‘DirectCXL’ can be found at CAMELab website (http://camelab.org/) and the video about accelerating the machine learning based recommendation model from Meta(Facebook) is available at CAMELab YouTube (https://youtu.be/jm8k-JM0qbM).