VMFlock: VM Co-Migration Appliance for the Cloud Samer Al-Kiswany With: Dinesh Subhraveti Prasenjit Sarkar Matei Ripeanu Why VM Migration across Data Centers ? For Cloud Users: • Freedom, Not locked to one cloud. • Private and public clouds. For the Cloud provider: • Load balancing across data centers • Accommodate scheduled maintenance • Arbitration of energy cost Requires: Efficient migration and fast instantiation of VM still images across data centers. 2 Challenges • Applications often deployed on multiple VMs (VM Flock) • Large VM image size (GBs) • Limited WAN bandwidth • Strict Cloud API • Limited resources (compute, memory, IO) 3 Opportunities • Similarities across VM flock images • Similarities across VM repositories • VM boot time access pattern 4 VMFlock Migration System An appliance for migrating and bootstrapping multiple VM images across data centers. Properties: • High performance • Scalable • Load balanced • Easy to adopt VMFlockMS achieves (compared to alternatives): • Up to 10x better compression rate • Up to 3.5x faster migration time 5 Outline • Introduction • VMFlockMS Design Migration Appliance Design VM Bootstrap component Design • Evaluation • Conclusion 6 VMFlockMS Design • Migration components Similarities across VM flock images Similarities across VM repositories • Bootstrap components VM boot pattern Source VM VMProfiler VM Repository Destination Migration nodes Migration nodes VM VMLaunchPad VM Repository 7 Migration Components Design Similarities across VM flock images Similarities across VM repositories Deduplication challenges: • CPU and IO intensive workload • Metadata size Source Destination Migration nodes VM Repository Migration nodes VM Repository 8 Migration Components Design Source Destination VM Repository 9 Migration Components Design Source Destination VM Repository 10 Migration Components Design Source Destination VM Repository 11 Migration Components Design Source VM Repository Destination VM Repository 12 Bootstrap Components Design Source Destination Migration nodes VM VMProfiler VM Repository Migration nodes VM VMLaunchPad VM Repository 13 Outline • Introduction • VMFlockMS Design Migration Appliance Design VM Bootstrap component Design • Evaluation • Conclusion 14 Evaluation – Setup Images • Application: Spree e-commerce (spree, sql, storage node) – 7.9 GB • Same-OS (Fedora: desktop, developer, server, plain) – 10.6 GB • Diff-OS (Fedora, openSUSE, Ubuntu) – 10.6 GB Alternatives • Gzip-All • Gzip-Separate (parallel Gzip) • Dedup-separate [Hirofuchi 09, Bradford 07, Sapuntzakis 02] 15 Evaluation – Setup Testbed • 2 machines at Almaden and 2 at T.J. Watson • 1 machine at each side works as a VM repository Courtesy Google Maps 16 Evaluation – Compression Rate Compression rate (%) 40 Diff-OS Same-OS App 35 35 30 25 20 15 10 30 25 20 15 10 5 5 0 0 Gzip All Gzip Separate Diff-OS Same-OS App 40 Compression rate (%) 45 Dedup Separate VMFlockMS No VM images at destination Gzip All Gzip Separate Dedup Separate VMFlockMS Single VM image at destination Achieves up to 10x better compression rate. 17 Evaluation – Migration Time 120 Migration time (min) 100 80 60 40 20 0 Gzip All Gzip Separate Dedup Separate VMFlockMS VMFlockMS VMFlockMS 1KB 8KB 128KB Migration time for the application flock VMFlockMS achieves up to 2x better migration time. 18 Evaluation – Migration Time 45 Migration time (min) 40 35 30 25 20 15 10 5 0 Gzip All Gzip Separate Dedup Separate VMFlockMS Migration time for the application flock using the VM repository emulator (emulating 4 SAS disks) VMFlockMS achieves up to 3.5x better migration time. 19 Evaluation – Boot Time • • • • VM Flock needs less then 20MB to boot. Most of the data already at destination VM flock boot in less than 1 min VMFlockMS overhead less than 3% 20 Summary VMFlockMS: An appliance for migrating and bootstrapping multiple VM images across data centers. Properties: • High performance • Scalable • Load balanced • Easy to adopt Where else: • Scientific data sets migration • Long execution pipelines 21 Thank you 22