Storage performance virtualization via throughput and latency control
ACM Transactions on Storage (TOS)
pClock: an arrival curve based approach for QoS guarantees in shared storage systems
Proceedings of the 2007 ACM SIGMETRICS international conference on Measurement and modeling of computer systems
mClock: handling throughput variability for hypervisor IO scheduling
OSDI'10 Proceedings of the 9th USENIX conference on Operating systems design and implementation
Dynamic resource allocation for spot markets in clouds
Hot-ICE'11 Proceedings of the 11th USENIX conference on Hot topics in management of internet, cloud, and enterprise networks and services
Nested QoS: providing flexible performance in shared IO environment
WIOV'11 Proceedings of the 3rd conference on I/O virtualization
Pesto: online storage performance management in virtualized datacenters
Proceedings of the 2nd ACM Symposium on Cloud Computing
Dynamic Resource Allocation for Spot Markets in Cloud Computing Environments
UCC '11 Proceedings of the 2011 Fourth IEEE International Conference on Utility and Cloud Computing
Hi-index | 0.00 |
Cloud-based services are emerging as an economical and convenient alternative for clients who don't want to acquire, maintain and operate their own IT equipment. Instead, customers purchase virtual machines (VMs) with certain Service Level Objectives (SLOs) to obtain computational resources. Existing algorithms for memory and CPU allocation are inadequate for I/O allocation, especially in clustered storage infrastructures where storage is distributed across multiple storage nodes. This paper focuses on: (1) dynamic SLO decomposition so VMs receive proper I/O service in each distributed storage node, and (2) efficient and robust local I/O scheduling strategy. To address these issues, we present pCloud, an adaptive I/O resource allocation algorithm that at runtime adjusts local SLOs. The local SLOs are generated for each VM at each storage node based on access patterns. We also adopt dual clocks in pCloud to allow automatic switching between two scheduling strategies. When system capacity is sufficient, pCloud interweaves requests in an earliest deadline first (EDF) manner. Otherwise resources are allocated proportionate to their normalized revenues. The results of our experiments suggest that pCloud is adaptive to various access patterns without significant manual pre-settings while maximizing profits.