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