A5018818656- Cloud Computing and Resource Management
- Parallel Computing and Optimization Techniques
- Advanced Data Storage Technologies
- Advanced Neural Network Applications
- Machine Learning and ELM
- Software-Defined Networks and 5G
- Security and Verification in Computing
- Neural Networks and Applications
Huawei Technologies (China)
2024
Peking University
2018-2024
Peng Cheng Laboratory
2018-2022
The tiered-memory system can effectively expand the memory capacity for virtual machines (VMs). However, virtualization introduces new challenges specifically in enforcing performance isolation, minimizing context switching, and providing resource overcommit. None of state-of-the-art designs consider address these challenges; we observe that a VM with tiered incurs up to 2× slowdown compared DRAM-only VM. We propose vTMM , hardware-software collaborative management framework virtualization....
The memory demand of virtual machines (VMs) is increasing, while the traditional DRAM-only system has limited capacity and high power consumption. tiered can effectively expand increase cost efficiency. Virtualization introduces new challenges for tiering, specifically enforcing performance isolation, minimizing context switching, providing resource overcommit. However, none state-of-the-art designs consider virtualization thus address these challenges; we observe that a VM with incurs up to...
Modern applications running on cloud data centers often consume a large amount of memory and their demands can vary during execution. Dynamic allocation is necessity for high utilization. For dataset application, using hugepages instead regular 4KB pages efficiently reduce access management overhead improve overall performance. Virtualization, which widely applied in server consolidation, brings new challenges to manage dynamically effectively, especially hugepages. In virtualized system,...
The overhead of memory virtualization remains nontrivial. traditional shadow paging (TSP) resorts to a page table (SPT) achieve the native walk speed, but updates require hypervisor interventions. Alternatively, nested enables low-overhead updates, utilizes hardware MMU perform long-latency two-dimensional walk. This paper proposes new solutions based on (machine) mode—the highest CPU privilege level in some architectures like Sunway and RISC-V. A programming interface, running mode,...
Virtualization is a key technique for supporting cloud services and memory virtualization major component of technology. Common mechanisms include shadow paging hardware-assisted paging. The model needs to synchronize shadow/guest page tables whenever there guest table update. In the design traditional (TSP), pages are write-protected so updates can be intercepted by hypervisor ensure synchronization. Frequent cause lots VM_Exits. Researchers have developed eliminate this overhead. However,...
With the rapid increase of data set size cloud and big applications, conventional regular 4KB pages can cause high pressure on hardware address translations. The becomes more prominent in a virtualized system, which adds an additional layer translation. Virtual to physical translations reply Translation Lookaside Buffer (TLB) cache mappings. However, even modern offers very limited number TLB entries. Meanwhile, misses significant performance degradation. Using 2MB or 1GB hugepages improve...
As more data-intensive tasks with large footprints are deployed in virtual machines (VMs), huge pages widely used to eliminate the increasing address translation overhead. However, once page mapping is established, all base regions share a single extended table (EPT) entry, so that hypervisor loses awareness of accesses regions. None state-of-the-art solutions can obtain access information at granularity for pages. We observe this lead incorrect decisions by hypervisor, such as data...
Neural networks have been widely applied to various research and production fields. However, most recent is focused on the establishment selection of a specific neural network model. Less attention paid their system overhead despite massive computing storage resource demand. This focuses relatively new direction that models system-level memory cache demand networks. We utilize learn predict hit ratio curve footprint with hyper-parameters as input. The prediction result used drive...