In current virtualized cloud platforms, resource provisioning strategy is still a big challenge. Provisioning will gain low resource utilization based on peak workload, and provisioning based on average work loads will sacrifice the potential revenue of cloud customers because of bad user experiences. VM-based performance isolation also restrains resource flowing on demand. As to memory, this eventually results in under-loaded memory and over-loaded memory in the same data center. This paper proposes a VM-oblivious dynamic memory optimization scheme, TMemCanal, which leverages under-loaded memory in a data center to accommodate the needs of loaded memory dynamically in a transparent fashion. TMemCanal is able to identify the under-loaded memory located in different VMs and reuse it in a way of memory flowing without any modification to their Guest OSs. We implemented TMemCanal through extending Xen hypervisor and evaluated using SpecWeb 2005 and LinkPack Benchmarks. Our evaluation shows that TMemCanal can efficiently save memory up to 50% with an overhead less than 7%. Our case study of server consolidation also shows TMemCanal can promote the performance of memory-intensive services up to 400%.
展开▼
