Bounding the Resource Savings of Utility Computing Models Artur Andrzejak, Martin Arlitt, Jerry Rolia Internet Systems and Storage Laboratory HP Laboratories Palo Alto HPL-2002-339 December 6th , 2002* E-mail: [email protected], [email protected], [email protected] utility computing, resource savings, grid computing In this paper we characterize resource usage in six data centers with approximately 1,000 servers. The results substantiate common knowledge that computing resources are typically under-utilized. Utility computing has been proposed as a way of increasing utilization and hence efficiency. Using an off-line integer programming model we bound the potential gain in efficiency for several different utility computing models. In a detailed study of a subset of the data center servers, we found that utility computing offered the potential to reduce the peak and mean number of CPUs needed by up to 53% and 79%, respectively. * Internal Accession Date Only Approved for External Publication  Copyright Hewlett-Packard Company 2002Bounding the Resource Savings of Utility Computing ModelsArtur Andrzejak∗Martin Arlitt Jerry RoliaInternet Systems and Storage LaboratoryHewlett Packard LaboratoriesPalo Alto, CA 94304Konrad-Zuse-Zentrum f¨ur Informationstechnik, Berlin (ZIB)Takustraße 7, 14195 Berlin-Dahlem, [email protected], [email protected], [email protected] 27, 2002AbstractIn this paper we characterize resource usage in six data centers with approximately 1,000 servers. Theresults substantiate common knowledge that computing resources are typically under-utilized. Utilitycomputing has been proposed as a way of increasing utilization and hence efficiency. Using an off-lineinteger programming model we bound the potential gain in efficiency for several different utility computingmodels. In a detailed study of a subset of the data center servers, we found that utility computing offeredthe potential to reduce the peak and mean number of CPUs needed by up to 53% and 79%, respectively.∗This work was done as Post-Doctoral research at Hewlett-Packard Laboratories.11 IntroductionData centers are large computing facilities with centralized resources and management infrastructure. Theyare often informally cited as having resource utilization rates of 20 to 30 percent. Utility computing has beenproposed as a sharing mechanism to make more effective use of computing infrastructure. Typically com-puting resources are allocated to individual (or a small set of) applications and provisioned for anticipatedpeak-load conditions over multi-year planning horizons. This can lead to sustained periods where resourcesare under-utilized. Sharing resources across multiple applications can: increase asset utilization by keepingresources busy, improve business agility by making available resources on demand, decrease power consump-tion by shifting work from lightly loaded resources – and placing them in a power savings mode, and loweroverall costs of ownership by reducing the quantity and space required for infrastructure and by increasingautomation.In this paper we study the load patterns of six data centers with approximately 1000 servers. We findthat resource utilization levels for 80% of the servers were indeed in the 30% range. A subset of servers forone data center is studied in detail.We describe several models of utility computing. Bounds for the potential resource savings of these utilitycomputing models are given for the subset of servers of the data center under detailed study. The bounds arebased on workload allocations found by an off-line integer programming model with a limited computationtime. The allocations are not guaranteed to be optimal, yet are closer to optimal than are likely to beachieved in practice with an on-line algorithm. For the servers under study we found the potential to reducethe peak and mean number of CPUs by up to 53 and 79%, respectively. These represent significant resourcesavings and opportunities for increasing asset utilization with respect to the capacity that would be requiredby data centers without resource sharing.The models of utility computing that we consider are summarized in Section 2. Section 3 discussesrelated work on utility computing. An analysis of loads for six data centers is presented in Section 4. Theoptimization methods used to bound the resource savings of the various models of utility computing aredescribed in Section 5. The results of the analysis are presented in Section 6. Section 7 gives summary andconcluding remarks.2 Utility Computing ModelsThere are many approaches towards utility computing. Broadly they fall into two categories. The first isa shared utility model where server resources are exploited by multiple customer applications at the sametime. The second is a more recent approach we define as a full server utility model where applicationsprogrammatically acquire and release servers as needed. A shared utility can act as an application to afull server utility. Inevitably both rely on network fabrics and possibly storage systems that are shared by2Horizontally scalable applicationDatabaseServerWeb Server 1App Server 1Web Server wApp Server a......Pools of serversof various capacitiesFull Server Utility Server 1 Full Server Utility Server p...AcquireRelease AcquireReleaseLoad 1, xShared Utility Server 1...Shared Server Compute Utility... Shared Utility Server mLoad n, yLoad n, 1 ...Load 1, 1 ... ...Full Server Compute UtilityCPU workload component (cpu x)Server 1 workload classData center workload (n servers)Figure 1: Shared and Full Server Models for Utility Computingmultiple applications. Figure 1 illustrates the relationships between these models.There are several major differences between the two approaches. Security and performance interactionwithin a server are concerns for the shared utility approach but not the full server approach, because eachserver is only allocated to one application at a time. Application control systems are needed to acquire andrelease resources at appropriate times for the full server utility approach. Whereas utility (server cluster)scheduling mechanisms aim to achieve the same effect for the shared approach without the need for applicationcontrol systems.There are many kinds of application workloads that must be served by such utilities. We distinguish theseas being either batch or business workloads. Batch