Power-Effi cient Management Infrastructure

The full awareness of VM execution context can be effectively used to take more effective DC reconfi guration decisions. The management infrastructure should exploit full visibility of any available information, namely, service/SLA/DC aware- ness , to guide VM placement decisions and runtime reconfi gurations. First, by using service types, the management infrastructure can better estimate side effects of VMs migration to schedule needed reconfi gurations and to decide the best mix of VMs for each host, in order to avoid placement solutions stressing too much a single- resource dimension only. Second, by using service providers’ SLAs , the management infrastructure can preventively validate Cloud reconfi gurations to increase placement stability ; also, it can issue reconfi guration commands to host hypervisors so to enforce resource reservation mechanisms. Finally, by using infor- mation on DC status, the management infrastructure can reduce the space of possi- ble reconfi guration decisions; for instance, number of powered hosts, current VM placement , and characterizations of VMs workloads are useful elements to guide DC reconfi gurations and to prevent expensive and useless migrations.

Putting together all these requirements, the management infrastructure of a power-effi cient Cloud has to address several heterogeneous aspects, including mon- itoring data gathering and processing, VM placement elaboration, and actuation of a suitable reconfi guration plans. To deal with such complexity, we can decompose the management infrastructure in three main stages as shown in Fig. 7.4 : the fi rst stage is in charge of DC monitoring, the second one exploits collected data to under- stand if a more suitable VM placement exists, and, fi nally, if that is the case, the third stage computes and applies a VM relocation plan . Hence, the proposed man- agement infrastructure consists of a Monitoring Layer , a Power-Effi cient Placement Computation Layer , and a DC Reconfi guration Layer that work together as a pipe- line; in the remainder, we better detail and clarify associated management duties of each layer.

The Monitoring Layer collects monitoring information associated with current DC confi guration. It gathers both host-level and DC-level information: host CPU, memory, and I/O load indicators are examples of host-level information; aggregated power consumption, network elements status, and cooling system state are, instead, examples of DC-level monitoring information. At the same time, this layer interacts

with deployed services to collect high-level monitoring data, such as average response times, depending on adopted service providers’ SLAs .

The Power-Effi cient Placement Computation Layer exploits information on service types, service providers’ SLAs , and DC status to assess if a better VM placement exists. Apart from monitoring information coming from the Monitoring Layer , it also queries external components of the management infrastructure to retrieve service profi les and service providers’ SLAs . Service profi les can be either supplied by service providers’ themselves or automatically detected by runtime profi ling and VM inspection. The former approach is simpler since service provid- ers will manually specify the attributes needed by the management infrastructure while the latter one requires active monitoring and introduces additional overhead.

When a better placement exists, the DC Reconfi guration Layer takes care of elaborating a suitable VM migration plan. Given both initial and target DC confi gu- rations, this layer is in charge of fi nding the minimum number of VM migrations to bring the DC to the target confi guration. All these VM migrations have to be arranged in a proper plan that will detail temporal dependencies between them; in brief, whenever possible this layer should try to execute more VM migrations in parallel to reduce the total reconfi guration time. Moreover, it is worth recalling that VM live migrations introduce both high computational overhead and high band- width consumption [ 4 ]. Hence, this layer has to exploit DC status to understand which migrations are possible and their possible consequences on running services. For instance, some migrations might be given higher priorities than others to free host resources required to accommodate incoming VMs; at the same time, multiple migrations can be executed in parallel when they do not lead to SLA violations.

To ground our design guidelines to a real Cloud management system imple- mentation, let us conclude this section by presenting some technical insights and seminal experimental results assessing the VM consolidation process in different service scenarios. In our previous work [ 34 ], we focused on the evaluation of

Monitoring Information Reconfiguration

Commands

Monitoring Layer Power Efficient Placement

Computation Layer DC Reconfiguration Layer

interferences between co-located VMs under different types of workloads; in particular, according to the taxonomy of Fig. 7.1 , we considered CPU-bound and

network-bound services . Here, we want to remark that VM consolidation is a very complex task since co-located VMs interfere among them in many different, and hard to predict, ways; that justifi es the adoption of service awareness to better drive the reconfi guration process. For the sake of brevity, we anticipate that all the results showed in this section are collected in a test bed made by two physical servers, each one with CPU Intel Core Duo E7600 @ 3.06 Ghz, 4 GB RAM, and 250 GB HD; also, we employed KVM as hypervisor, and all VMs have 1 VCPU and 512 MB RAM.

In the fi rst set of experiments, we considered VMs executing CPU-bound services that introduce a VCPU load in {20, 40, 60, 80, and 99 %}, as shown in Fig. 7.5 . We considered different consolidation ratios, spanning from 1 to 5 VMs, on the same host; since the physical host of these experiments has a dual-core CPU, the aggregate CPU consumption can reach 200 %. Moving to higher consolidation ratios, we experience performance degradation when the aggregated VCPU con- sumption is higher than available CPU; however, as reported in Fig. 7.5 , we are always able to exploit all available resources, by reaching 200 % CPU utilization, meaning that the two CPU cores are fully loaded. Moreover, our experiments demonstrate (not reported here for the sake of brevity) that VMs get a fair share of available resources, and the performance degradation is easily predictable by considering the overcommit ratio of CPU resources.

In our second set of experiments, instead, we considered VMs running an Apache Web (AB) server that supplies static HTML pages as an example of network-bound service. In this case, we exploited the standard AB tool to load the Web servers [ 35 ]; Fig. 7.5 Interferences between co-located VMs under CPU-bound services

different AB clients, one for each VM, emit 100,000 requests with a number of concurrent requests in {2, 4, 6, 8}. By analyzing the experimental results reported in Fig. 7.6 , we can see that, with consolidation ratios higher than 2, the hypervisor is not able to grant all the available resources. We think that behavior is probably due to the fact that different VMs try to concurrently access the same network inter- face, thus leading to additional synchronization bottlenecks that prevent the full exploitation of all the available CPU resources.

In conclusion, as stated before, the Power-Effi cient Placement Computation Layer should exploit service awareness to better evaluate possible side effects and interfer- ences between co-located VMs. In addition, SLA awareness is important to prevent service violations, but it must be considered that SLA degradation for network- bound services is not always easily predictable as in the case of CPU- bound services.

7.6

Conclusion

Power-effi cient management infrastructures for Cloud computing are gaining increased attention in the last years due to both environmental and economical issues. Modern DCs can introduce power consumption in the order of MWatts; hence, they require intelligent mechanisms to selectively switch off not required hosts and network elements, so to increase fi nal power effi ciency . Power-effi cient Cloud infrastructure management has two positive effects: fi rst, it reduces the envi- ronmental impact of the DC; second, it increases the economical revenue of the Cloud provider due to lower DC operational costs. At the same time, by using two Fig. 7.6 Interferences between co-located VMs under network-bound services

signifi cant case studies, i.e., CPU-bound and network-bound services, we have shown that VM performance degradation due to VM consolidation cannot be easily predicted and depends on many factors, including virtualization support and type of services executed in the VM. Although different works are currently available in this research area, we think that additional research is required to handle both host and networking aspects at the same time.

First of all, since hosts represent the most important source of DC power con- sumption, it is fundamental to effi ciently place the current workload; however, according to some studies [ 36 ], networking can account for the 5–10 % of the DC power consumption, and that justifi es the usage of software-defi ned network tech- niques to further reduce total DC power consumption. At the same time, since both service and SLA awareness can be very useful to quickly identify and prioritize feasible reconfi guration decisions (as shown by our preliminarily experimental results), we think that additional work is needed also to better exploit them in order to dominate the complexity of such placement problems. Finally, it is important to devise new power optimization formulations to consider in the decision process not only single VMs but also groups of multiple VMs with high and correlated traffi c, such as in the case of VMs collaborating to the same data processing task.

References

1. Lenk, A.: What’s inside the Cloud? An architectural map of the Cloud landscape. In: Proceedings of ICSE Workshop on Software Engineering Challenges of Cloud Computing. IEEE Computer Society, pp. 23–31 (2009)

2. Armbrust, M. et al.: Above the Clouds: A Berkeley view of Cloud Computing. EECS Department, University of California, Berkeley (2009)

3. Buyya, R., et al.: Cloud computing and emerging IT platforms: vision, hype, and reality for delivering computing as the 5th utility. Futur. Gener. Comput. Syst. 25 , 599–616 (2009) 4. Clark, C. et al.: Live migration of virtual machines. In: Proceedings of the 2nd Symposium on

Networked Systems Design & Implementation, 2005

5. Le, K. et al.: Reducing electricity cost through virtual machine placement in high performance computing clouds. In: Proceedings of the Conference on High Performance Computing, Networking, Storage and Analysis (SC ’11), 2011

6. Pettey, C.: Gartner estimates ICT industry accounts for 2 percent of global CO 2 emissions.

http://www.gartner.com/it/page.jsp?id=503867 (2007). Accessed 6 Oct 2012

7. Murugesan, S.: Harnessing green IT: principles and practices. IEEE IT Prof. 10 , 24–33 (2008) 8. Baliga, J.: Green Cloud Computing: balancing energy in processing, storage, and transport.

IEEE J. 99 , 149–167 (2011)

9. Beloglazov, A. et al.: A taxonomy and survey of energy-effi cient data centers and Cloud Computing systems. In: Proceedings of CoRR, 2010

10. Heller, B. et al.: ElasticTree: saving energy in data center networks. In: Proceedings of the 7th USENIX Conference on Networked Systems Design and Implementation (NSDI’10), 2010 11. Biran, O. et al.: A stable network-aware VM placement for Cloud Systems. In: Proceedings of

the IEEE CCGrid’12, Ottawa, 2012

12. Meng, X. et al.: Improving the scalability of data center networks with traffi c-aware virtual machine placement. In: Proceedings of the 29th Conference on Information Communications (INFOCOM’10), 2010

14. Elson, J., Howell, J.: Handling fl ash crowds from your garage. In: Proceedings of the USENIX 2008 Annual Technical Conference on Annual Technical Conference (ATC’08). USENIX, Berkeley (2008)

15. Voorsluys, W. et al.: Cost of virtual machine live migration in Clouds: a performance evaluation. Proceedings of the 1st International Conference on Cloud Computing (CloudCom’09), 2009 16. Breitgand, D. et al.: Cost-aware live migration of services in the Cloud. In: Proceedings of the

3rd Annual Haifa Experimental Systems Conference (SYSTOR ’10), 2010

17. Kayacan, E.: Grey system theory-based models in time series prediction. Expert Syst. Appl. 37 , 1784–1789 (2010)

18. Isci, C. et al.: Runtime demand estimation for effective dynamic resource management. In: Proceedings of the IEEE Network Operations and Management Symposium (NOMS), 2010 19. Al-Fares, M. et al.: A scalable, commodity data center network architecture. In: Proceedings

of the ACM SIGCOMM 2008 Conference on Data communication (SIGCOMM ’08), 2008 20. Greenberg, A. et al.: VL2: a scalable and fl exible data center network. In: Proceedings of the

ACM SIGCOMM 2009 Conference on Data communication (SIGCOMM ’09), 2009 21. Guo, C. et al.: BCube: a high performance, server-centric network architecture for modular data

centers. In: Proceedings of SIGCOMM Computer Communication. ACM, pp. 63–74 (2009) 22. Verma, A. et al.: pMapper: power and migration cost aware application placement in virtual-

ized systems. In: Proceedings of the 9th ACM/IFIP/USENIX International Conference on Middleware (Middleware ’08), 2008

23. Kim, K.H. et al.: Power-aware provisioning of Cloud resources for real-time services, Proc. 7th International Workshop on Middleware for Grids, Clouds and e-Science (MGC 2009), 2009 24. Abdelsalam, H.S. et al.: Analysis of energy effi ciency in clouds. In: Proceedings, Computation

World: Future Computing, Service Computation, Cognitive, Adaptive, Content, Patterns, 2009 25. Beloglazov, A., Buyya, R.: Energy effi cient resource management in virtualized Cloud Data Centers. In: Proceedings of the 10th IEEE/ACM International Conference on Cluster, Cloud and Grid Computing, 2010

26. Buyya, R. et al.: Modeling and simulation of scalable cloud computing environments and the CloudSim toolkit: challenges and opportunities. In: Proceedings of the 7th High Performance Computing and Simulation Conference (HPCS’09), 2009

27. Van, H.N. et al.: Performance and power management for Cloud infrastructures. In: Proceedings of the IEEE 3rd International Conference on Cloud Computing (CLOUD’10), 2010

28. Jung, G. et al.: Mistral: dynamically managing power, performance, and adaptation cost in Cloud infrastructures. In: Proceedings of the IEEE 30th International Conference on Distributed Computing Systems (ICDCS’10), 2010

29. Wang, Y., Wang, X.: Power optimization with performance assurance for multi-tier applica- tions in virtualized data centers. In: Proceedings of the 39th International Conference on Parallel Processing Workshops (ICPPW ’10), 2010

30. Dupont, C. et al.: An energy aware framework for virtual machine placement in cloud feder- ated data centres. In: Proceedings of the 3rd International Conference on Future Energy Systems: Where Energy, Computing and Communication Meet (e-Energy ’12), 2012 31. Mann, V. et al.: VMFlow: leveraging VM mobility to reduce network power costs in data

centers. In: Proceedings of the IFIP Networking, 2011

32. Korupolu, M. et al.: Coupled placement in modern data centers. In: Proceedings of the IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2009

33. Wang, M. et al.: Consolidating virtual machines with dynamic bandwidth demand in data centers. In: Proceedings of the IEEE INFOCOM 2011 Mini-Conference, 2011

34. Corradi, A. et al.: VM consolidation: a real case based on OpenStack Cloud. Future Gener. Comput. Syst., Elsevier Science, SI on Management of Cloud Systems, Available online Jun. 2012, doi: 10.1016/j.future.2012.05.012 . pp. 1–10 (2012)

35. Kvitka, C.: ab – Apache HTTP server benchmarking tool. http://httpd.apache.org/docs/2.0/ programs/ab.html . Accessed 8 Apr 2013

36. Hamilton, J.: Energy proportional datacenter networks. http://perspectives.mvdirona. com/2010/08/01/EnergyProportionalDatacenterNetworks.aspx (2010). Accessed 6 Oct 2012