Presentation is loading. Please wait.

Presentation is loading. Please wait.

A workload-aware energy model for VM migration

Similar presentations


Presentation on theme: "A workload-aware energy model for VM migration"— Presentation transcript:

1 A workload-aware energy model for VM migration
Vincenzo De Maio, Gabor Kecskemeti, Radu Prodan Distributed and Parallel Systems University of Innsbruck

2 Summary Introduction Motivation Background Model Design
Experimental setup Evaluation Conclusion and future work

3 Motivation Energy efficiency has become a serious issue for virtual data centres; Around 30-40B$ in 2007. Physical machines (PMs) are often under-utilised; 10-50% utilisation for most of the time. IDEA: improving energy efficiency of data centres by increasing PM’s utilization. Workload consolidation.

4 Workload consolidation
Mapping VMs on a reduced subset of PMs. PM 1 VM 1 (30%) PM 2 VM 2 (30%) PM 3 VM 3 (30%) PM 1 VM 1 (30%) VM 2 (30%) VM 3 (30%) PM 2 PM 3

5 Consolidation: pros and cons
Increases utilisation of PMs; Save energy by shutting down the unused PMs. CONs: Drawbacks in performance are possible; A bad VM-PM mapping may increase energy consumption. Necessity of predicting energy efficiency of a VM-PM mapping.

6 Evaluating consolidation energy-efficiency
Assessing energy efficiency of VM-PM mappings: Energy prediction models; Such models should consider each data centre aspect; Startup/Shutdown of PMs, resource utilization… VM migration

7 VM Migration We distinguish between: Non-live migration Live migration
The capability of moving the state of a VM from one PM (source PM) to another PM (target PM) without losing consistency. We distinguish between: Non-live migration Suspend/resume Live migration Pre-copy/suspend/resume

8 VM migration energy consumption
Usually not considered for consolidation; Existing VM migration models do not consider: Impact of actors’s workload on VM migration; Source/Target PM, Migrating VM, … Overcommitment. There is significant room for improvement in the existing models.

9 Our contribution We propose WAVM3, a model for VM migration considering: Actors’ workload; Overcommitment; We target scenarios with homogeneous hosts; Based on Xen’s limitations.

10 Summary Introduction Motivation Background Model Design
Experimental setup Evaluation Conclusion and future work

11 VM migration actors

12 Identifying main actors
Migrating VM Source host Target host Consolidation manager Issues the migration, but then it does not further affects it External network load Constant

13 Workload types CPU-BOUND Computationally-intensive applications.
MEMORY-BOUND Using more than 90% of memory. Memory pages are continuously updated. Dirtying rate: percentage of memory pages dirtied at each iteration

14 Workload impact non-live migration
If source/target hosts are fully loaded CPU contention Longer state transfer Higher energy consumption for high CPU utilisation. Step 3 Resumed Step 2 State transferred Step 1 Suspended

15 Workload impact live migration
Same problems of non-live migrations for CPU-BOUND workloads Plus the continuous memory updates More iterations Higher energy consumption for high dirtying rate Step 5 Resumed Step 4 Suspended Step 3 Repeat step 2 Step 2 Target updated Step 1 State moved while running

16 Workload impact Summary
Migration type Migrating VM Source host Target host CPU-BOUND NON-LIVE Depending on Source/Target Load Slowdown for initiation and state transfer Slowdown for activation and state transfer LIVE MEMORY-BOUND No influence Slight performance degradation Multiple transfer of VM state

17 Migration energy phases
Normal execution phase Each actor performs its normal operation. Initiation phase From the issuing of the migration to its start. Transfer phase The state of the VM is transferred over the network. Service activation phase The VM is started on the target and the source releases the resources previously owned by the VM.

18 WAVM3 Energy model Global energy model:
𝑚 𝑠 is the instant when the migration starts 𝑚 𝑒 depends on VM dirtying rate, VM size and hosts’ load.

19 WAVM3 Energy model 𝑚 𝑒 = 𝑇 𝑖𝑛𝑖𝑡 (𝑣,ℎ)+ 𝑇 𝑡𝑟𝑎𝑛𝑠𝑓 (𝑣,ℎ)+ 𝑇 𝑎𝑐𝑡𝑖𝑣 (𝑣,ℎ)
𝑚 𝑒 = 𝑇 𝑖𝑛𝑖𝑡 (𝑣,ℎ)+ 𝑇 𝑡𝑟𝑎𝑛𝑠𝑓 (𝑣,ℎ)+ 𝑇 𝑎𝑐𝑡𝑖𝑣 (𝑣,ℎ) 𝑇 𝑖𝑛𝑖𝑡 𝑣,ℎ = 𝑇 𝑎𝑐𝑡𝑖𝑣 𝑣,ℎ = 𝑅𝐴𝑀(𝑣) 𝐵𝑊 𝑖𝑜 (ℎ) 𝑇 𝑡𝑟𝑎𝑛𝑠𝑓 𝑣,ℎ = 𝑅𝐴𝑀(𝑣) 𝐵𝑊 𝑛𝑒𝑡 (ℎ,𝑡) + 𝑖=1 𝐼 𝐷𝑃(𝑣,𝑖) 𝐵𝑊 𝑛𝑒𝑡 (ℎ,𝑡) 𝐼: number of copy iterations. 𝐷𝑃 𝑣,𝑖 : Dirty pages at iteration i.

20 WAVM3 Power model Initiation Phase Transfer Phase Activation Phase

21 Experimental setup VM instances
INSTANCE NAME CPU [%] MEMORY [%] DIRTYING RATE [%] VMCPU 100 5 VMMEM [5-95]

22 Experimental setup Experiments
Source host configuration Target host configuration Migrating VM CPULOAD-SOURCE Increasing CPU load IDLE VMCPU CPULOAD-TARGET MEMLOAD-VM VMMEM VMMEM [5-95]% DR MEMLOAD-SOURCE VMMEM 95% DR MEMLOAD-TARGET

23 Measurement infrastructure

24 Migration energy phases (NON-LIVE)

25 Migration energy phases (LIVE)

26 WAVM3 evaluation PM ERROR [%] (non-live) (live) 2 x AMD Opteron 8356
2 x Intel Xeon E5-2690 SOURCE 11.8 12.5 12.7 TARGET 12 16.3 5 17.2

27 Model comparison MODEL HOST ERROR [%] (non-live) (live) WAVM3 SOURCE
11.8 TARGET 12 5 CPU only (Huang et al., 2012) 15.7 12.8 12.9 NET+MEM (Liu et al., 2011) 26.9 36.3 25.3 29.4 NET only (Strunk et al., 2013) 17.7 35.4 30 36.2

28 Conclusion and future work
We increased accuracy of existing models up to 24% (vs NET+MEM) Future work: Studying the impact of network-intensive workloads On different types of network (10G, Infiniband…) Considering swapping Integrating our model in a Cloud simulator

29 Thanks for your attention! Q & A


Download ppt "A workload-aware energy model for VM migration"

Similar presentations


Ads by Google