Acryl (CEO Park Oe-jin), an artificial intelligence (AI) company, announced on the 14th that its AAAI research team presented new possibilities for remote direct memory access (RDMA) technology through the latest paper 'PeRF: Preemption-enabled RDMA FRAMEwork'.
RDMA is one of the technologies for transmitting data in a computer network. Network communication usually involves the CPU intervening in the middle to copy and transmit data when sending or receiving data, but with RDMA, the network adapter directly accesses the memory to copy and transmit data. This saves processor resources required for data transmission and reduces transmission delays.
Acryl presented this paper at the 'USENIX Annual Technical Conference' held in Santa Clara, California, USA from the 10th to the 12th. 'USENIX ATC' is an academic conference recognized as the most prestigious in the fields of system software and network research. It was founded in 1975 to study the Unix OS series and related systems. At the time of its establishment, it was called the 'Unix Users Group' and changed its name to the current USENIX in June 1977. It has been publishing a technical journal called 'Login' since its establishment. Its headquarters is at the University of California, Berkeley.
Acryl emphasized that this paper's announcement "will be an opportunity for our company's research achievements to be recognized internationally." Existing RDMA technology was optimized for a single-tenant environment, making it difficult to solve performance isolation, security, and scalability issues in a multi-tenant cloud environment. Acryl's PeRF (Preemption-enabled RDMA FRAMEwork) was designed to overcome these limitations. It is a new RDMA framework that provides software-based performance isolation.
Dr. Lee Soo-ki of Acryl explains the ReRF paper that suggests new possibilities for RDMA technology.
Tenant is a cloud term that refers to an entity that rents another building rather than its own. In other words, it uses the cloud resources of the service provider, not its own resources. 'Multi-tenancy', which means multiple tenancies, refers to dividing a cloud resource and providing it to multiple users, just like renting out a house by dividing it. It is done for cost efficiency by having multiple tenants in a single resource. Tenants in multi-tenancy receive fast updates and upgrades from the service provider.
Acryl emphasized about PeRF, "It dynamically controls the use of RDMA resources by each tenant by utilizing an innovative RNIC (RDMA NIC) preemption mechanism. This allows flexible performance isolation while maintaining the original performance of RDMA, and shows better performance than existing methods." "In particular, PeRF is optimized for data-intensive applications, so it can be used in various fields such as big data analysis, machine learning, distributed storage, and key-value storage."
■ Maximizing GPU Utilization: Providing Cost-Effective Solutions
Digital Camp Advertisement Logo
With the advent of the generational AI era, as GPU prices have skyrocketed, economically utilizing GPUs has become an issue, and PeRF can make a big contribution to solving this problem, Acryl explained. In other words, PeRF enables efficient distribution and optimization of GPU resources through RDMA technology, which means that users can achieve maximum performance even with a small number of GPUs. This reduces the burden of purchasing expensive GPUs and maximizes the utilization of existing GPU infrastructure, Acryl said.
■ Maximizing Performance in Combination with Multi-path RDMA
In particular, 'PeRF' achieves greater synergy effects when combined with the software-based 'Multi-path RDMA' technology developed by Acryl. 'Multi-path RDMA' technology maximizes network efficiency and reliability by distributing and transmitting data through multiple paths. Acryl said that in particular, when combined with 'PeRF', it further improves performance isolation and resource utilization in a multi-tenant environment.
Acryl emphasized, "This combination is especially useful when multiple tenants use RDMA simultaneously in cloud-based data centers. PeRF monitors each tenant's RDMA usage in real time and adjusts resource allocation as needed to ensure performance isolation. At the same time, Multi-path RDMA optimizes the data transmission path to maximize network efficiency." He added, "The combination of these two technologies will dramatically improve data transmission speed and stability in cloud environments."
■ AI Platform and PeRF: MLOps and LLMOps Performance Enhancements
Acryl applied PeRF to its flagship product, Jonathan, an MLOps and LLMOps platform, to further enhance the performance and stability of the product. The company said, "Customers using Jonathan can build an optimal machine learning model operation environment," and "They can also experience excellent performance and efficiency in large-scale language model operation."
■ PeRF's Outstanding Performance Proven in Real Tests
Acryl announced that ReRF showed far superior performance compared to existing performance isolation technologies in performance test results. This test was conducted by a joint research team led by Dr. Yeom Ik-jun of Sungkyunkwan University, and participated by Dr. Lee Soo-ki and Researcher Choi Min-gyu of Acryl, and Professor Kim Young-hoon of Sungkyunkwan University. The research team announced that the test results showed that PeRF maximizes RDMA performance in a multi-tenant cloud environment while providing flexible performance isolation, and that it is expected to be utilized in various data-intensive applications in the future.
■ LLMOps Platform Research Vision and Future Plans
Park Oe-jin, CEO of Acryl, said, “This announcement will allow many people in academia and industry to confirm the potential of PeRF and Jonathan,” and added, “Acryl is a representative AI company in Korea, and is striving to create an AI infrastructure that is easy for anyone to use and economical.” He added, “Acryl is leading the AI industry through various innovative technologies and is committed to providing the best AI solutions to customers.”