Improving Agility and Elasticity in Bare-metal Clouds Yushi Omote † , Takahiro Shinagawa ‡ , Kazuhiko Kato † † University of Tsukuba, ‡ The University of Tokyo 1
Bare-metal Clouds An IaaS for high performance and device functionality OS transparent User No Virtual OS Machine Physical Machine Provider 2
OS-deployment Problem Long wait time sacrifices agility and elasticity (1) Image Copy (2) Reboot from Local Disk (Tens of minutes) (A few minutes) User OS OS Installer User OS Image Image Image Server 3
Existing Approach 1 OS Streaming Deployment [Clerc et al. IPCCC’10] Network Boot + Background Copy User OS Agility and Elasticity Performance Special Driver Image OS-specific drivers Server are required. ✗ OS transparency 4
Existing Approach 2 Conventional VMMs [VMware’01, Xen’03, KVM’07] Streaming deployment with VMMs User OS Agility and Elasticity OS transparency VMM Continuous Image Server virtualization overhead ✗ Performance 5
OS Deployment with a Special-purpose VMM 1) Streaming deployment 2) Seamless de-virtualization Agility and Elasticity Performance OS transparency User OS User OS VMM VMM 6
Challenge Expose & Control Physical Devices Virtual Direct I/O? Devices? ✗ OS Control I/Os VMM ✗ Expose physical interface 7
Device-interface-level I/O mediation A device mediator performs: (1) I/O interpretation OS Device Driver to understand I/O context VMM (2) I/O redirection Device to perform network booting Mediator (3) I/O multiplexing to perform background install Physical device interface 8
＝ I/O Interpretation Determine when/how to mediate I/O requests OS Device Driver VMM Understand state transitions Device based on monitoring I/O Mediator Device State Transitions 9
I/O Redirection LBA=4 OS Data NUM=8 (1) Interpret LBA=4 (2) Redirect Interrupt NUM=8 Image Small Server Request VMM (3) Restart Disk 10
I/O Multiplexing Status OS OS Check Request (1) Request Idle VMM State Request (2) Emulate (3) Queue Image Server VMM Disk 11
CPU/Memory Virtualization for De-virtualizable VMM Memory CPU OS Guest Physical Address = VMM VMM Physical Address Identity Mapping No indirection VMM exposes physical memory VMM runs passively with VMX Mark VMM regions as reserved No guest scheduling (via BIOS INT15/e802) 12
De-virtualization (3) Turns off (2) Turns off (1) Turns off IO CPU virtualization nested paging VM exits OS OS OS Device Driver VMM VMM VMM H/W Ease Find safe I/O timing Unsynchronized VM exits condition TLB flush (VMXOFF Issue) 13
Performance Evaluation • Deployed 32-GB OS Image (Ubuntu 14.04 64-bit) • OS-startup Time A HPC Cluster • Cassandra Throughput • Storage Throughput Intel Xeon X5680 (3.33 GHz) / 96GB RAM • InfiniBand Latency HDD 500GB/7200 RPM SATA Mellanox InfiniBand (4X QDR) Intel 82575 EM GbE Network Card Interconnected by A Mellanox Grid Director InfiniBand Switch & A FUJITSU SR- S348TC1 GbE Switch 14
OS-startup Time Image Copy Reboot+Firminit. VMM Boot OS Boot 370 145 29 Image Copy Proposed Quick start up 5+58 (8.6 times faster) OS Streaming 49 (NFSRoot) VM Streaming 30+42 (KVM/NFS) 0.00 150.00 300.00 450.00 600.00 Elapsed Time (sec) 15
Cassandra Throughput (Throughout Deployment) Proposed KVM (No Background Install) Eventual 120% bare-metal 110% % of Baremetal performance 100% 90% Seamless 80% de-virtualization 70% 50 200 350 500 650 800 950 1100 1250 Elapsed Time (sec) 16
Storage Throughput Read Write Bare-metal performance 120.00 117 115 Throughput (MB/sec) 112 112 112 112 101 100 90.00 60.00 30.00 0.00 Bare-metal Deploy Devirt KVM/Local 17
InfiniBand RDMA latency 1.70 1.61 Bare-metal performance 1.28 1.30 1.30 1.30 Latency (usec) 0.85 0.43 0.00 Baremetal Deploy Devirt KVM/Pass 18
Conclusion • Improved agility and elasticity in bare-metal clouds • De-virtualizable VMM with streaming deployment Device-interface-level I/O mediation • • Achieved quick startup of an OS 8.6 times faster than image copy • Preserved high performance & OS-transparency • 19
Future work • Generating device mediators from specification • Reduce development cost of device mediators • More advanced features of IaaS clouds • Live migration and checkpointing 20
Thank you 21
Recommend
More recommend
