1/6 Paper Reference Number: WCEEENG_A00_000 The effect of machine virtualization on the environmental impact of desktop environments Dr. Mahmoud Abaza1, Daryl Allenby1,2 1 Athabasca University, 2 The Northern Alberta Institute of Technology (NAIT), Canada II. REQUIREMENTS ANALYSIS T Abstract-This paper evaluates two important and recently prominent ideas in computing science in terms of their combined potential to provide responsive and environmentally responsible alternatives to traditional desktop computing deployments. The ideas of virtualization and green computing are combined to develop a production model for virtual desktop environments. This virtual model is then studied to determine possible concurrencies when varying desktop operating system and secondary storage input/output performance variables. The potential for concurrency in the production model is based on the use of modern equipment, hypervisor processing hosts, Virtual Machines (VMs) and thin clients as end user devices. The virtual model, as a final presentation, is then compared to equal numbers of traditionally deployed desktop computers. Keywords – Virtual Machines, green computing, hypervisor, environment friendly. I. INTRODUCTION The model being studied is comprised of VMs hosted on servers running VMware’s ESX Server version 3.5 and thin clients. We refer to this coupling of VM and thin client as Virtual Desktop Infrastructure (VDI). The research will determine the number of VMs that can be supported by a given amount of processing power as it varies by Operating System (OS) and secondary storage. The limit of concurrency will be equal to the maximum useable number of VMs as determined by each variable pair. Each VM will employ a reasonable analogue of actual user activity achieved by scripted programmatic tasks performed in the VM. This VDI model is then be evaluated as to its potential to achieve a higher degree of efficiency in terms of power, cooling and ultimately CO2 produced when compared to an identical number of traditionally deployed machines. The results of the evaluation can be interpreted in terms of net reduction in the total cost of energy needed to supply both solutions for a period of years. The total cost of energy is determined by the pounds of CO2 contributed by each solution. VM host OS and secondary storage are expected to be important variables because they are likely to become bottlenecks sooner than other components of the system [1], [2]. Varying OS type under the included selections will expose efficiencies in software architecture. Varying secondary storage presentations under the included selections will show how decreasingly available Input/Output Operations per Second (IOpS) affects the upper limit of concurrency. Secondary storage or Logical Unit Number (LUN) types and OS types chosen have been done so to reflect likely choices that organizations are apt to make if implementing a VDI based solution [8]. III. BENCHMARKING AND DATA COLLECTION The purpose of benchmarking will be to determine empirically that the limit of concurrency under each variable has been reached. The study will simulate user activity on each VM. The simulations are achieved by a series of host shell scripts written that incorporate several typical user activates. For example each shell script opens a word processing document, an email client, a spread sheet, copy a large file, delete a large file, open and close several different web sites that are specifically chosen for their dynamic content and displayed media types, play a video and open/close a PDF document. Each activity is punctuated by an appropriate intervening time delay. The entire sequence runs for approximately twenty minutes and loop infinitely. Between the Microsoft operating systems the identical shell scripts, source files and productivity software can be studied. The Ubuntu desktop VMs will be unique in both shell script implementation, productivity software, and will have some differences in source file types. The shell scripts are designed to produce an acceptable randomness of hypervisor resource consumption as realized as an average over time. After the template VMs are created, n additional VMs will be spawned from their templates, the processing, memory access and I/O read write patterns of the group will become randomized as their preconfigured shell scripts: start, close, launch and use applications. This will negate most of the benefit realized by caching that the processing host uses at the processor, memory and storage controller. The 2009 World Congress on Electronics and Electrical Engineering, WCEEENG'09 April 6-9, 2009, Cairo, Egypt 2/6 Paper Reference Number: WCEEENG_A00_000 IV. THE CONCURRENT LIMIT When adding additional VMs to determine the concurrent limit for each pair of variables; that limit will be met when one of two things happens after the addition of an active VM to the hypervisor: 1. A critical hypervisor resource reaches (as a percentage averaged over one hour) 75 percent, in scope resources include CPU cycles (in MHz), memory (in KB), network throughput (in KBps) and disk IOpS (in KBps) ; 2. The user experience in VM n + 1 (assessed qualitatively) becomes unacceptable in terms of responsiveness. The goal in stopping at 75 percent average utilization is to allow for unexpected user activity in production, as a useable production model is an important outcome of the study. At 100 percent utilization of any one resource more VM concurrence would be possible, but the user experience would be poor and not a realistic implementation. A good production model will allow for a reasonable and achievable statement of concurrency. Succinctly that; a given server configuration produces x number of VMs; this plus thin clients equal to concurrent running VMs requires x amount of kWh and produces n BTUs. Compare that to an equal number of standard desktops requiring x’ kWh and producing n’ BTUs. The second condition represents a safeguard; if for example, after adding one more VM to the server configuration, the response time in terms of general usability is unacceptable; then concurrency will be considered to be that number less one. This would be true even if the critical hypervisor resources were all within the 75 percent limit. This condition covers any influencing variables that are difficult or impossible to quantify. One of the expected unquantifiable variables that might be encountered is that of instantaneous latency. Latency, in this context, is the momentary delay in the delivery of data, over a bus or network, which is not reflected well in measures of throughput or, in our case, IOpS. This latency may affect VM performance in ways that would otherwise be difficult to measure or predict, but still manifest as a user experience that is unacceptable. The qualitative evaluation will be simple; if the first item in the user simulating shell script does not launch instantly, the experiment enters a caution phase. If, after a delay of five minutes, the same delay is observed, the concurrency limit will be considered met. If, after the delay, the observation is not repeated, it will exit the caution phase and VM additions continue normally. Hypervisor resource utilization data will be collected from the ESX management interface regarding processor, memory and I/O requirements of each VM and the total committed by the hypervisor. V. POWER CONSUMPTION MONITORING Once the limit of concurrency has been determined under each variable pair, measurement of amps drawn at concurrency will be measured by inline electrical meters. All equipment involved in the study will be connected to instrumented circuits. The thin client and its LCD panel power usage will be determined, while in active use, under a RDP session to one of the VMs. This number will be multiplied by the concurrent maximum determined at the hypervisor and added to the total energy consumed by the virtual model. To determine the energy used by the equivalent traditional deployment of desktop computers, the same operating systems, applications, source files and shell scripts will be used, as was done in the VMs. All three operating systems will be tested on a single desktop configuration, and the energy consumed each time will be multiplied by the concurrent maximum determined at the hypervisor. VI. CALCULATIONS AND FORMULA After finding the sum in watts for each model, the determination of net kilograms CO2 can be made, and those calculations extrapolated in years to determine the total cost of energy for each. During calculations the following assumptions will be used: 1. There is 0.613 kilograms of CO2 in every kWh [3]. This figure is based on the average across all forms of generation according to the source; 2. Each kWh used produces 3,412 BTU of heat [5]; 3. The instruments measuring amps at the circuit breaker have a manufacture’s stated accuracy of ±0.0375 amps. The results of calculations will therefore have an uncertainty of at least ±0.0375 and possibly more if the results of multiple circuits are added. In no case, will the results of more than five circuits be combined so we can confidently state an uncertainty in measurement of ±.1875 watts for all experiments. The primary formulas used to determine net pounds of CO 2 are described below and flow from one to another in descending order: 1. kWh × cost in CO2 per kWh = kilograms of CO2 contributed by processing; 2. Kilograms of CO2 contributed by processing + kilograms of CO2 contributes by cooling = net kilograms of CO2. VII. VDI EXPERIMENTAL RESULTS Table 2, shows the results of extrapolation and compare both solutions is terms of carbon measuring a single day’s operation. TABLE 2: CARBON COST AFTER EXTRAPOLATION The 2009 World Congress on Electronics and Electrical Engineering, WCEEENG'09 April 6-9, 2009, Cairo, Egypt 3/6 Paper Reference Number: WCEEENG_A00_000 VDI Power (kWh) Carbon Cost (kg) Desktop Power (kWh) Carbon Cost (kg) Carbon Delta Ubuntu FC-SCSI 1345.99 825.09 1527.59 936.41 -111.32 Ubuntu FATA 1272.60 780.10 1527.59 936.41 -156.31 Ubuntu iSCSI 1316.25 806.86 1527.59 936.41 -129.55 Ubuntu NFS 1388.27 851.01 1527.59 936.41 -85.40 Ubuntu SAS 1287.33 789.14 1527.59 936.41 -147.27 XP SP2 FC-SCSI 1433.50 878.74 1568.71 961.62 -82.88 XP SP2 FATA 1376.37 843.71 1568.71 961.62 -117.90 XP SP2 iSCSI 1405.07 861.31 1568.71 961.62 -100.31 XP SP2 NFS 1445.10 885.85 1568.71 961.62 -75.77 XP SP2 SAS 1380.70 846.37 1568.71 961.62 -115.25 Vista SP1 FCSCSI 1587.13 972.91 1631.69 1000.23 -27.31 Vista SP1 FATA 1476.60 905.16 1631.69 1000.23 -95.07 Vista SP1 iSCSI 1540.88 944.56 1631.69 1000.23 -55.66 Vista SP1 NFS 1554.43 952.87 1631.69 1000.23 -47.36 Vista SP1 SAS 1489.33 912.96 1631.69 1000.23 -87.26 In looking closely at our data we can make some general conclusions and recommendations: 1. VDI must be scaled up to hundreds of VMs before it starts to have an environmental advantage over standard desktop deployments; 2. Once properly scaled up, VDI, in all the forms tested, is a more environmentally friendly solution than standard desktop deployments by an average of 9.98 percent or 95.64 kilograms of CO2 saved for every day, for a deployment of approximately 750 stations. This number becomes even more impressive when extended to the year over year saving at 34,908.60 kilograms of CO 2 saved ; 3. FATA based VMs are the top environmental performers, regardless of installed VM OS, by an average of 12.82 percent or 123.09 kilograms of CO 2 saved every day for a deployment of approximately 750 stations. Once again, this number is more impressive when extended year over year at 44,927.85 kilograms of CO2 saved; 4. Choice of operating system does make a difference; we see that Ubuntu scores the best, with XP SP2 needing on average 18.18 percent more resources to match the same number of VMs and Vista SP1 needing on average 41.78 percent more resources to reach an equal number. We include a delta column where positive values (if present) indicate a solution that is more costly to the environment. Values near zero show that the solutions are competitive and, finally, negative numbers indicated solutions that are less costly to the environment. We now have all the information we need to make some conclusive statements about VDI’s performance in terms of greenness. Acknowledgment VIII. CONCLUSIONS AND RECOMMENDATIONS Figure 2, shows summary results in terms of carbon performance, sorted by the delta between VDI and traditional desktops after one year of operation. This work was supported by NAIT in the provision of all computing and instrumentation equipment necessary to perform the experiments and collect data. For additional information or questions on the study please contact Daryl Allenby by email at daryl.allenby@gmail.com. FIGURE 2: SUMMARY PERFORMANCE IN CARBON SORTED BY DELTA References [1] Ahmad, Irfan, et al. 2003. An Analysis of Disk Performance in VMware ESX Server Virtual Machines. VMware Corporation Web Site. [Online] October 27, 2003. [Cited: February 8, 2008.] http://www.vmware.com/pdf/wwc_performance.pdf. The 2009 World Congress on Electronics and Electrical Engineering, WCEEENG'09 April 6-9, 2009, Cairo, Egypt Paper Reference Number: WCEEENG_A00_000 [2] Buytaert, Kris, et al. 2007. Best Damn Server Virtualization Book Period. Burlington : Syngress Publishing, Inc., 2007. [3] Energy, Department of and Agency, Environmental Protection. 1999. Carbon Dioxide Emissions from the Generation of Electric Power in the United States. Environmental Protection Agency. [Online] October 15, 1999. [Cited: January 30, 2008.] http://yosemite.epa.gov/oar/globalwarming.nsf/UniqueKey Lookup/SHSU5BPL2K/$File/co2emiss99.pdf. [4] Energy, U.S. Department of. 2007. Electric Power Annual . Energy Information Administration. [Online] October 22, 2007. [Cited: July 6 , 2008.] http://www.eia.doe.gov/cneaf/electricity/epa/epa.pdf. [5] Larabie, Chuck. 2003. Power and storage: the hidden cost of ownership - Storage Management. Business Technology Network. [Online] October 2003. [Cited: February 18, 2008.] The 2009 World Congress on Electronics and Electrical Engineering, WCEEENG'09 April 6-9, 2009, Cairo, Egypt 4/6