MANAGEMENT OF HETEROGENEOUS WORKLOADS
Systems and methods for managing a system of heterogeneous workloads are provided. Work that enters the system is separated into a plurality of heterogeneous workloads. A plurality of high-level quality of service goals is gathered. At least one of the plurality of high-level quality of service goals corresponds to each of the plurality of heterogeneous workloads. A plurality of control functions are determined that are provided by virtualizations on one or more containers in which one or more of the plurality of heterogeneous workloads run. An expected utility of a plurality of settings of at least one of the plurality of control functions is determined in response to the plurality of high-level quality of service goals. At least one of the plurality of control functions is exercised in response to the expected utility to effect changes in the behavior of the system.
Latest IBM Patents:
- Lateral silicon nanospikes fabricated using metal-assisted chemical etching
- Waveguide architecture for photonic neural component
- Beacon placement and distribution
- Use of simultaneously received videos by a system to generate a quality of experience value
- Determining hardware requirements for a wireless network event using crowdsourcing
CROSS-REFERENCE TO RELATED APPLICATION
This application is a divisional of U.S. patent application Ser. No. 11/741,875 filed on Apr. 30, 2007, the disclosure of which is incorporated by reference herein in its entirety.
FIELD OF THE INVENTION
The present invention relates generally to application management, and more particularly, to methods and apparatus for management of heterogeneous workloads
BACKGROUND OF THE INVENTION
Many organizations rely on a heterogeneous set of applications to deliver critical services to their customers and partners. This set of applications includes web workloads typically hosted on a collection of clustered application servers and a back-end tier database. The application mix also includes non-interactive workloads such as portfolio analysis, document indexing, and various types of scientific computations. To efficiently utilize the computing power of their datacenters, organizations allow these heterogeneous workloads to execute on the same set of hardware resources and need a resource management technology to determine the most effective allocation of resources to particular workloads.
A traditional approach to resource management for heterogeneous workloads is to configure resource allocation policies that govern the division of computing power among web and non-interactive workloads based on temporal or resource utilization conditions. With a temporal policy, the resource reservation for web workloads varies between peak and off-peak hours. Resource utilization policies allow non-interactive workload to be executed when resource consumption by web workload falls below a certain threshold. Typically, resource allocation is performed with a granularity of a full server machine, as it is difficult to configure and enforce policies that allow server machines to be shared among workloads. Coarse-grained resource management based on temporal or resource utilization policies has previously been automated. See, K. Appleby et al., “Oceano—SLA-Based Management of a Computing Utility,” IFIP/IEEE Symposium on Integrated Network Management, Seattle, Wash., May 2001; and Y. Hamadi, “Continuous Resources Allocation in Internet Data Centers,” IEEE/ACM International Symposium on Cluster Computing and the Grid, Cardiff, UK, May 2005, pp. 566-573.
Once server machines are assigned to either the web or the non-interactive workload, existing resource management policies can be used to manage individual web and non-interactive applications. In the case of web workloads, these management techniques involve flow control and dynamic application placement. See, C. Li et al., “Performance Guarantees for Cluster-Based Internet Services,” IEEE/ACM International Symposium on Cluster Computing and the Grid, Tokyo, Japan, May 2003; G. Pacifici et al., “Performance Management for Cluster-Based Web Services,” IEEE Journal on Selected Areas in Communications, Vol. 23, No. 12, December 2005; and A. Karve et al., “Dynamic Placement for Clustered Web Applications,” World Wide Web Conference, Edinburgh, Scotland, May 2006. In the case of non-interactive workloads, the techniques involve job scheduling, which may be performed based on various existing scheduling disciplines. See, D. Feitelson et al., “Parallel Job Scheduling—a Status Report,” 10th Workshop on Job Scheduling Strategies for Parallel Processing, 2004, pp. 1-16. To effectively manage heterogeneous workloads, a solution is needed that combines flow control and dynamic placement techniques with job scheduling.
SUMMARY OF THE INVENTION
The embodiments of present invention provide a system and method for management of heterogeneous workloads.
For example, in one aspect of the present invention, a method for managing a system of heterogeneous workloads is provided. Work that enters the system is separated into a plurality of heterogeneous workloads. A plurality of high-level quality of service goals is gathered. At least one of the plurality of high-level quality of service goals corresponds to each of the plurality of heterogeneous workloads. A plurality of control functions are determined that are provided by virtualizations on one or more containers in which one or more of the plurality of heterogeneous workloads run. An expected utility of a plurality of settings of at least one of the plurality of control functions is determined in response to the plurality of high-level quality of service goals. At least one of the plurality of control functions is exercised in response to the expected utility to effect changes in the behavior of the system.
In additional embodiments of the present invention, Web applications are placed on one or more of a plurality of heterogeneous server machines through a placement controller driven by utility functions of allocated CPU demand. Web application requests are received at a request router. The web application requests are dispatched from the request router to one or more web applications on one or more of the plurality of heterogeneous server machines in accordance with a scheduling mechanism. The scheduling mechanism is dynamically adjusted in response to at least one of workload intensity and system configuration. Jobs are allocated to one or more of the plurality of heterogeneous server machines in accordance with placement decisions communicated to a job scheduler by the placement controller.
In another aspect of the present invention, a system for management of heterogeneous workloads comprises a plurality of heterogeneous server machines. The system further comprises a placement controller, driven by utility functions of allocated computer processing unit (CPU) demand, which places web applications on one or more of the plurality of heterogeneous server machines. A request router receives and dispatches requests to one or more web applications on one or more of the plurality of heterogeneous server machines in accordance with a scheduling mechanism. A flow controller in communication with the request router and the placement controller dynamically adjusts the scheduling mechanism in response to at least one of workload intensity and system configuration. A job scheduler allocates jobs to one or more of the plurality of heterogeneous server machines in accordance with placement decisions communicated to the job scheduler by the placement controller.
In additional embodiments of the present invention, the system may also comprise a web application workload profiler that obtains profiles for web application requests in the form of an average number of CPU cycles consumed by requests of a given flow, and provides the profiles for web application requests to the flow controller and placement controller. The system may also comprise a job workload profiler that obtains profiles for jobs in the form of at least one of the number of CPU cycles required to complete the job, the number of threads used by the job, and the maximum CPU speed at which the job may progress, and provides the profiles for jobs to the job scheduler.
In further embodiments of the present invention the system may comprise a plurality of domains in the each of the plurality of heterogeneous server machines. A first domain of the plurality of domains may comprise a node agent in communication with the placement controller and the job scheduler. A second domain of the plurality of domains may comprise a machine agent in communication with the node agent that manages virtual machines inside a given heterogeneous server machine. The node agent provides job management functionality within the heterogeneous server machine through interaction with the machine agent. The machine agent is capable of at least one of creating and configures a virtual machine image for a new domain, copying files from the second domain to another domain, starting a process in another domain, and controlling the mapping of physical resources to virtual resources.
A system is provided that considerably improves the way heterogeneous workloads are managed on a set of heterogeneous server machines using automation mechanisms provided by server virtualization technologies. The system introduces several novel features. First, it allows heterogeneous workloads to be collocated on any server machine, thus reducing the granularity of resource allocation. This is an important aspect for many organizations that rely on a small set of powerful machines to deliver their services, as it allows for a more effective resource allocation when any workload requires a fractional machine allocation to meet its goals. Second, the approach uses high-level performance goals, as opposed to lower-level resource requirements, to drive resource allocation. Hence, unlike previous techniques, which manage virtual machines according to their defined resource requirements, an embodiment of the present invention provides an application-centric view of the system in which a virtual machine is only a tool used to achieve performance objectives. Third, an embodiment of the present invention exploits a range of new automation mechanisms that will also benefit a system with a homogeneous, particularly non-interactive, workload by allowing more effective scheduling of jobs.
BRIEF DESCRIPTION OF THE DRAWINGS
DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS
Integrated automated management of heterogeneous workloads is a challenging problem for several reasons. First, performance goals for different workloads tend to be of different types. For interactive workloads, goals are typically defined in terms of average or percentile response time or throughput over a certain time interval, while performance goals for non-interactive workloads concern the performance of individual jobs. Second, the time scale of management is different. Due to the nature of their performance goals and short duration of individual requests, interactive workloads lend themselves to automation at short control cycles. Non-interactive workloads typically require calculation of a schedule for an extended period of time. Extending the time scale of management requires long-term forecasting of workload intensity and job arrivals, which is a difficult if not impossible problem to solve. Server virtualization assists in avoiding this issue by providing automation mechanisms by which resource allocation may be continuously adjusted to the changing environment. Thus, to collocate applications on a physical resource, one must know the applications' behavior with respect to resource usage and be able to enforce a particular resource allocation decision. For web applications, with the help of an L7 gateway, one can rather easily observe workload characteristics and, taking advantage of similarity of web requests and their large number, derive reasonably accurate short-time predictions regarding the behavior of future requests. Non-interactive jobs do not exhibit the same self-similarity and abundance properties, hence predicting their behavior is much harder. Enforcing a resource allocation decision for web workloads can also be achieved relatively easily by using flow control mechanism. Server virtualization gives us similar enforcement mechanisms for non-interactive applications.
While server virtualization allows for better management of workloads to their respective SLA goals, it also introduces considerable challenges in order to use it effectively. They concern the configuration and maintenance of virtual images, infrastructure requirements to make an effective use of the available automation mechanisms, and the development of algorithmic techniques capable of utilizing the larger number of degrees of freedom introduced by virtualization technologies. Embodiments of the present invention address some of these challenges.
Referring initially to
Referring now to
The management architecture of
In the system considered, overload protection for interactive workloads is provided by an L7 request router 212 which implements a flow control technique. Router 212 classifies incoming requests into flows depending on their target application and service class, and places them in per-flow queues. Requests are dispatched from the queues based on weighted-fair scheduling discipline, which observes a system-wide concurrency limit. The concurrency limit ensures that all the flows combined do not use more than their allocated re-source share. The weights further divide the allocated resource share among applications and flows.
Both the concurrency limit and scheduling weights are dynamically adjusted by a flow controller 216 in response to changing workload intensity and system configuration. Flow controller 216 builds a model of the system that allows it to predict the performance of the flow for any choice of concurrency limit and weights via optimizer 218. This model may also be used to predict workload performance for a particular allocation of CPU power. The functionality of flow controller 216 is used to come up with a utility function for each web application at utility function calculator 220, which gives a measure of application happiness with a particular allocation of CPU power given its current workload intensity and performance goal.
Long-running jobs are submitted to the system via job scheduler 214, which, unlike traditional schedulers, does not make job execution and placement decisions. In the system, job scheduler 214 only manages dependencies among jobs and performs resource matchmaking Once dependencies are resolved and a set of eligible nodes is determined, jobs are submitted to an application placement controller (APC) 222 via a job queue manager 224.
Each job has an associated performance goal. An embodiment of the present invention supports completion time goals, but the system may be extended to handle other performance objectives. From this completion time goal an objective function is derived which is a function of actual job completion time. When job completes exactly on schedule, the value of the objective function is zero. Otherwise, the value increases or decreases linearly depending on the distance of completion time from the goal.
Job scheduler 214 uses APC 222 as an adviser to where and when a job should be executed. When APC 222 makes a placement decision, actions pertaining to long-running jobs are returned to job scheduler 214 and put into effect via a job executor component 226. Job executor 226 monitors job status and makes it available to APC 222 for use in subsequent control cycles.
APC 222 provides the decision-making logic that affects placement of both web and non-interactive workloads. To learn about jobs in the system and their current status, APC 222 interacts with job scheduler 214 via a job scheduler proxy 228. A placement optimizer 230 calculates the placement that maximizes the minimum utility across all applications. It is able to allocate CPU and memory to applications based on their CPU and memory requirements, where memory requirement of an application instance is assumed not to depend on the intensity of workload that reaches the instance. The optimization algorithm of APC 222 is improved; its inputs are modified from application CPU demand to a per-application utility function of allocated CPU speed, and the optimization objective is changed from maximizing the total satisfied CPU demand to maximizing the minimum utility across all applications. A web application placement executor 232 places applications on nodes 202, 204, 206 in an optimized manner.
Since APC 222 is driven by utility functions of allocated CPU demand and, for non-interactive workloads, objective functions of achieved completion times are only given, a way to map completion time into CPU demand, and vice versa, may also be provided. Recall that for web traffic a similar mechanism exists, provided by the flow controller. The required mapping is very difficult to obtain for non-interactive workloads, because the performance of a given job is not independent of CPU allocation to other jobs. After all, when not all jobs can simultaneously run in the system, the completion time of a job that is waiting in the queue for other jobs to complete before it may be started depends on how quickly the jobs that were started ahead of it complete, hence it depends on the CPU allocation to other jobs. In the system, simple but effective heuristics are implemented that allow aggregate CPU requirements to be estimated for all long-running jobs for a given value of utility function at job utility estimator 234. This estimation is used to obtain a set of data-points from which the utility function is later extrapolated. This estimation is used to obtain a set of data-points from which values needed to solve the optimization problem are later extrapolated.
To manage web and non-interactive workloads, APC relies on the knowledge of resource consumption by individual requests and jobs. The system includes profilers for both kinds of workloads. A web workload profiler 236 obtains profiles for web requests in the form of the average number of CPU cycles consumed by requests of a given flow. A job workload profiler 238 obtains profiles for jobs in the form of the number of CPU cycles required to complete the job, the number of threads used by the job, and the maximum CPU speed at which the job may progress.
Features of virtualization of which the system in an embodiment of the present invention is capable of taking advantage are briefly enumerated.
- PAUSE When a virtual machine is paused, it does not receive any time on the node's processors, but the virtual machine remains in memory.
- RESUME Resumption is the opposite of pausing—the virtual machine is once again allocated execution time on the node's processors.
- SUSPEND When a virtual machine is suspended, its memory image is saved to disk, and it is unloaded.
- RESTORE Restoration is the opposition of suspension—an image of the virtual machine's memory is loaded from disk, and the virtual machine is permitted to run again.
- MIGRATE Migration is the process of moving a virtual machine from one node to another. In standard migration, the virtual machine is first paused, then the memory image is transferred across the network to the target node, and the virtual machine is resumed.
- LIVE_MIGRATE Live migration is a version of migration in which the virtual machine is paused. Instead, the memory image is transferred over the network whilst the virtual machine runs, and (when the memory images on both nodes match) control passes to the new host.
- MOVE_AND_RESTORE When a virtual machine has already been suspended, and needs to be restored on a different node, the management system must first move the saved memory image to the new node, and then restore the virtual machine on the new host node.
- RESOURCE_CONTROL Resource control modifies the amounts of various resources that virtual machines can consume, such as, for example, CPU and memory.
While virtualization can be provided using various technologies, an embodiment of the present invention uses Xen as it is capable of providing the wide variety of controls discussed above. Xen is an x86 virtual machine monitor capable of running multiple commodity operating systems on shared hardware. Although it typically requires that guest operating systems be modified, user-level code can execute in guest VMS, called domains, without modification. Xen provides a series of controls, including those discussed above. All of these controls are most directly accessible from a special domain on each Xen-enabled node, labeled domain 0.
The system relies on an entry gateway that provides flow control for web requests. The entry gateway provides a type of high-level virtualization for web requests by dividing CPU capacity of managed nodes among competing flows. Together with an overload protection mechanism, the entry gateway facilitates performance isolation for web applications.
Server virtualization could also be used to provide performance isolation for web applications. This would come with a memory overhead caused by additional copies of the OS that would have to be present on the node. Hence, it is believed that middleware virtualization technology is a better choice for managing the performance of web workloads.
Since middleware virtualization technology can only work for applications whose request-flow it can control, a lower level mechanism must be used to provide performance isolation for other types of applications. As outlined in the previous section, server virtualization provides powerful mechanisms to control resource allocation of non-web applications.
Referring now to
Xen is used to provide on-line automation for resource management, hence it is desirable to make management actions light-weight and efficient. This consideration concerns the process of creating virtual images, which may be quite time consuming. Substantial delays are avoided, which would otherwise be incurred each time it is intended to start a job from job scheduler 314, by pre-creating a set of images in accordance with a command executor 304 and a repository 306, for use during runtime. The dispensing of these pre-created images is performed by image management subsystem 308. Images once used to run a process are scrubbed of that process data and may be reused by future processes. In small-scale testing thus far, it has been found sufficient to pre-create a small number of images; however, image management subsystem 308 may be extended to dynamically extend the pool of available images if needed.
Inside a created image, a new process may be created. This is done by populating the image with the files necessary to run that new process. In the system, it is assumed that the files required for all processes that may run on the node are placed in its domain 0 in advance. Hence, there is only a need to copy them from domain 0 to the created image. Clearly, there are mechanisms that would allow us to transfer files from an external repository to a node where the process is intended to run.
Before it may be booted, an image must be provided with configuration files to set up its devices and networking. This functionality is encapsulated by a configuration management subsystem 310. To assign an IP address and DNS name, a DHCP server can be used, although in the system a simpler, more restrictive, module has been implemented that selects configuration settings from a pool of available values.
Referring now to
Referring again to
CPU allocation to domains may be controlled in Xen using three mechanisms. First, the number of virtual CPUs (vCPUs) can be selected for any VM. Typically, the number of vCPUs depends on the parallelism level of a process that executes inside a domain. Second, vCPUs may be mapped to physical CPUs at a virtual-to-physical resource mapper 312. By ‘pinning’ vCPUs of a domain to different physical CPUs the performance of the domain may be improved. Finally, CPU time slices may be configured for each domain. When all vCPUs of a domain are mapped to different physical CPUs, allocation of 50 out of 100 time slices to the domain implies that each vCPU of the domain will receive 50% of the compute power of the physical CPU to which it is mapped. Xen also permits borrowing, by which CPU slices allocated to a domain that does not need them can instead be used by other domains.
In a default configuration provided by Xen, each domain receives the same number of vCPUs as there are physical CPUs on a machine. Each of those vCPUs will be mapped to a different physical CPU and receives 0 time slices with CPU borrowing turned on. In the process of managing the system this allocation is modified inside virtual-to-physical resource mapper 312. When a domain is first started, Xen is allowed to create the default number of vCPUs and map them to different physical CPUs. Only the number of time slices is set to obtain the CPU allocation requested by placement controller. While domain is running, its actual CPU usage is observed. If it turns out that the domain is not able to utilize all vCPUs it has been given, it may be concluded that the job is not multi-threaded. Hence, to receive its allocated CPU share, its vCPUs must be appropriately reduced and remapped. Virtual-to-physical resource mapper 312 must attempt to find a mapping that provides the domain with the required amount of CPU power spread across the number of vCPUs that the job in the domain can use—clearly, this is not always possible.
All the VM actions provided by the machine agent are asynchronous JMX calls. They are followed by JMX notifications indicating the completion of an action.
To hide the usage of VMs from a user, a higher-layer of abstraction is implemented, embedded inside a node agent 316, which provides the job management functionality. It provides operations to start, pause, resume, suspend, restore, and resource control a job. To implement these operations, the node agent interacts with command executor 304 of machine agent 302 in domain 0 using its VM management interfaces. When a job is first started, node agent 310 creates (or obtains a pre-created) image in which to run the job. It records the mapping between the job ID and VM ID in a VM ID repository 318. Then it asks machine agent 302 to copy corresponding process binaries to the new image and to boot the image. Once domain is running, the job is started inside it.
Observe that a job is placed in its own domain. This provides performance isolation among jobs such that their individual resource usage is controlled, but it comes at the expense of added memory overhead. The system maybe extended such that it allows collocation of multiple jobs inside a single domain based on some policies.
The node agent process is placed in domain 1, which is the domain used for all web applications. There are two reasons for placing the node agent in a separate domain than domain 0. First, the application server middleware already provides a node agent process with all required management support, thus adding new functionality is a matter of a simple plug-in. Second, domain 0 is intended to remain small and light-weight. Hence, using it to run functionality that does not directly invoke VM management tools is avoided.
Like the machine agent, the node agent exposes its API using JMX.
Referring now to
In block 518, dependencies among jobs are managed at a job scheduler. In block 520, jobs are matched with nodes. In block 522, jobs are received from the job scheduler at a job scheduler proxy of the placement controller. In block 524, requirements for jobs for a given value of utility function are estimated at a job utility estimator of the placement controller. In block 526, job placements are calculated that maximizes the minimum utility across all web applications at a placement optimizer of the placement controller. In block 528, the job placements are communicated to the job scheduler through the job scheduler proxy. In block 530, jobs are allocated to the nodes in accordance with placement decisions communicated to a job scheduler by the placement controller.
In block 532, profiles for web application request are obtained in the form of an average number of CPU cycles consumed by requests of a given flow. In block 534, profiles for jobs are obtained in the form of CPU cycles required to complete the job, the number of threads used by the job or the maximum CPU speed at which the job may progress. Finally, in block 536, the profiles for web application request are communicated to the placement controller and the flow controller, and the profiles for jobs are communicated to the job scheduler.
Provided below are examples of experimental evaluations, implementations and integrations, additional implementations for embodiments of the present invention are also possible.
An approach of the present invention is experimentally evaluated using both real system measurements and a simulation. The system in an embodiment of the present invention has been implemented and integrated with WebSphere Extended Deployment application server middleware. WebSphere Extended Deployment is used to provide flow control for web applications and use Xen virtual machines to provide performance isolation for non-interactive workloads.
In the experiments, a single micro-benchmark web application is used that performs some CPU intensive calculation interleaved with sleep times, which simulate backend database access or I/O operations. A set of non-interactive applications is also used, which consists of well known CPU-intensive benchmarks. In particular, BLAST, from The National Center for Biotechnology Information (NCBI), Lucene, from Apache Software Foundation, ImageMagick, and POV-Ray, from Persistence of Vision Pty. Ltd., are used as representative applications for bioinformatics, document indexing, image processing and 3D rendering scenarios respectively. BLAST (Basic Local Alignment Search Tool) is a set of similarity search programs designed to explore all of the available sequence databases for protein or DNA queries. Apache Lucene is a high-performance, full-featured, open-source text search engine library written entirely in Java. In the experiments, the example indexing application provided has been run with the Lucene library to index a large set of files previously deployed in the file system. POV-ray (Persistence of Vision Raytracer) is a high-quality free tool for creating three-dimensional graphics. ImageMagick is a software suite to create, edit, and compose bitmap images.
Referring now to
The system in an embodiment of the present invention is experimented on a cluster of two physical machines, xd018 and xd020, each with two 3 GHz CPUs and 2 GB memory. The XenSource-provided Xen 3.0.2 packages are used for RedHat Enterprise Linux 4.
While testing, it is determined that the resource control actions of the version of Xen are rather brittle and cause various internal failures across the entire Xen machine. Therefore, in the experiments, resource control actions in the machine agent code are suppressed.
The effectiveness of automation mechanisms used by the system in an embodiment of the present invention may be studied. Three different jobs are taken from the set, JOB1, JOB2, and JOB5, and perform various automation actions on them while measuring their duration. Migration is not measured because it is not set up in the system in an embodiment of the present invention. Instead, move-and-restore is used. Clearly, this is quite an inefficient process, mostly due to the overhead of copying the image. A dramatically different result is expected once live-migration is put in place.
Referring now to
An experiment to demonstrate the benefits of using server virtualization technology in the management of heterogeneous workloads is described. StockTrade (a web-based transactional test application) is deployed in domain on two machines xd018 and xd020. Load to StockTrade is varied using a workload generator that allows for control of the number of client sessions that reach an application. Initially, 55 sessions are started and with this load it is observed that response time of StockTrade requests is about 380 ms and approaches response time goal of 500 ms. Referring now to
At 800 s since the beginning of the experiment, load is reduced to StockTrade to 25 concurrent client sessions. When CPU usage of StockTrade reduces to about 50% of each machine, the placement controller decides (E) to start JOB1 (F) on xd018. After 1000 s, the number of client sessions is increased back to 55; placement controller suspends JOB1 (G). Typically, JOB1 will later be resumed when any of the following conditions occur: (1) JOB5 completes, (2) load to StockTrade is reduced, or (3) JOB1 gets close enough to its target completion time so as to necessitate its resumption, even at the expense of worsened performance for StockTrade. However, the occurrence of the third condition indicates that the system in an embodiment of the present invention is under-provisioned; hence SLA violation may not be avoided. This simple experiment demonstrates that with the use of server virtualization, the system in an embodiment of the present invention is able to balance resource usage between web and non-interactive workloads.
The usefulness of server virtualization technology is shown in the management of homogeneous, in this case non-interactive workloads. Using the same experimental set-up as in Section V-B, a test case is run that involves only long-running jobs shown in
The test case is started by submitting JOB1 (A), which is started on xd020 and takes its entire CPU power. Soon after JOB1 is submitted, JOB2 and JOB3 (B) are submitted, which both get started on xd018 and each of them is allocated one CPU on the machine. Ten minutes later, JOB4 (C) is submitted, which has a very strict completion time requirement. In order to meet this requirement, APC decides to suspend JOB1 and start JOB4 in its place. Note that if JOB1 was allowed to complete before JOB4 is allowed to start, JOB4 would wait 5 min in the queue, hence it would complete no earlier than 13 min after its submission time, which would exceed its goal. Instead, JOB4 is started as soon as it arrives and completes within 10 min, which is within its goal. While JOB4 is running, JOB5 (D) is submitted. However, JOB5 belongs to a lower class than any job currently running, and therefore is placed in the queue. When JOB4 completes, JOB5 is started on xd020. Since JOB5 consumes only 1 CPU, APC also resumes JOB1 and allocates it the remaining CPU. However, to avoid Xen stability problems in the presence of resource control mechanisms, the resource control action is suppressed. As a result, resolving competition for CPUs is delegated to Xen hypervisor.
In the next phase of the experiment, the use of migration is demonstrated. After the completion of JOB1 and JOB3, submit JOB6 (E) is submitted. When JOB6 arrives, JOB2 and JOB5 each consume 1 CPU on xd018 and xd020 respectively. Since JOB6 requires 2 CPUs, APC may either (1) make it wait in the queue, (2) suspend JOB2 or JOB5, (3) collocate and resource control JOB6 with either JOB2 or JOB5, or (4) migrate either JOB2 or JOB5. Options (1)-(3) would result in wasted capacity on one or both machines. Moreover, options (1) and (3) would result in having platinum class job receive proportionately less CPU power than JOB5, whose service class is gold. This would clearly not be the optimal decision from the perspective of the optimization objective. Hence, APC decides (E) to move JOB4 to xd018 (which it will now share with JOB5) and start JOB6 on the now-empty xd020.
Even though this experiment shows that APC correctly uses migration when machine fragmentation makes it difficult to place new jobs, it also demonstrates a limitation of the optimization technique, which is currently oblivious to the cost of performing automation actions. Although in this experiment, 15 min is an acceptable price to pay for migrating a job, it is easy to imagine a scenario, where performing such a costly migration would have a damaging effect.
Potential benefits of using virtualization technology in the management of non-interactive workloads are studied. A system is simulated in which jobs with characteristics similar to the ones in Table I are submitted randomly with exponentially distributed inter-arrival times. The workload mix includes 25% multithreaded jobs with execution time of 32 min, 25% multithreaded jobs with execution time of 23 min, 25% single-threaded jobs with execution time of 66 min, 15% single-threaded jobs with execution time of 45 min, and 10% single-threaded jobs with execution time of 127 min. The service class distribution for all jobs is 50%, 40%, and 10% for platinum, gold, and silver service class, respectively. Mean inter-arrival time is varied between 8 and 30 min.
The simulation does not model the cost of performing virtualization actions. Hence, the results concern the theoretical bound on the particular algorithmic technique used.
The placement algorithm (APC) is evaluated with well known scheduling techniques: fast-come-fast-serve (FCFS) and earliest-deadline-first (EDF), in which completion time goal is interpreted as deadline. The placement technique is executed after disabling automation mechanisms provided by virtualization technology (APC_NO_KNOBS).
Referring now to
Referring now to
Referring now to
As shown, the computer system may be implemented in accordance with a processor 1310, a memory 1312, I/O devices 1314, and a network interface 1316, coupled via a computer bus 1318 or alternate connection arrangement.
It is to be appreciated that the term “processor” as used herein is intended to include any processing device, such as, for example, one that includes a CPU (central processing unit) and/or other processing circuitry. It is also to be understood that the term “processor” may refer to more than one processing device and that various elements associated with a processing device may be shared by other processing devices.
The term “memory” as used herein is intended to include memory associated with a processor or CPU, such as, for example, RAM, ROM, a fixed memory device (e.g., hard drive), a removable memory device (e.g., diskette), flash memory, etc.
In addition, the phrase “input/output devices” or “I/O devices” as used herein is intended to include, for example, one or more input devices for entering speech or text into the processing unit, and/or one or more output devices for outputting speech associated with the processing unit. The user input speech and the speech-to-speech translation system output speech may be provided in accordance with one or more of the I/O devices.
Still further, the phrase “network interface” as used herein is intended to include, for example, one or more transceivers to permit the computer system to communicate with another computer system via an appropriate communications protocol.
Software components including instructions or code for performing the methodologies described herein may be stored in one or more of the associated memory devices (e.g., ROM, fixed or removable memory) and, when ready to be utilized, loaded in part or in whole (e.g., into RAM) and executed by a CPU.
While previous techniques concentrate on managing virtual machines as primary abstractions that are exposed to end user, the technique manages applications using automation mechanisms provided by virtual servers. An application-centric approach is taken and the usage of VM technology is kept as invisible to end user as possible.
The system allows management of heterogeneous workloads on a set of heterogeneous server machines using automation mechanisms provided by server virtualization technologies. The system introduces several novel features. First, it allows heterogeneous workloads to be collocated on any server machine, thus reducing the granularity of resource allocation. Second, the approach uses high-level performance goals (as opposed to lower-level resource requirements) to drive resource allocation. Third, the technique exploits a range of new automation mechanisms that will also benefit a system with a homogeneous, particularly non-interactive, workload by allowing more effective scheduling of jobs.
Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be made by one skilled in the art without departing from the scope or spirit of the invention.
1. A system for management of heterogeneous workloads comprising:
- a plurality of heterogeneous server machines;
- a placement controller, driven by utility functions of allocated computer processing unit (CPU) demand, that places web applications on one or more of the plurality of heterogeneous server machines;
- a request router that receives and dispatches requests to one or more web applications on one or more of the plurality of heterogeneous server machines in accordance with a scheduling mechanism;
- a flow controller in communication with the request router and the placement controller that dynamically adjusts the scheduling mechanism in response to at least one of workload intensity and system configuration; and
- a job scheduler that allocates jobs to one or more of the plurality of heterogeneous server machines in accordance with placement decisions communicated to the job scheduler by the placement controller.
2. The system of claim 1, wherein the web applications placed on one or more of the plurality of heterogeneous server machines comprise application server clusters.
3. The system of claim 1, wherein the placement controller comprises:
- a job scheduler proxy that receives jobs from the job scheduler and communicates job placement decisions to the job scheduler;
- a job utility estimator that estimates requirements for jobs for a given value of utility function;
- a placement optimizer that calculates job and web application placement that maximizes the minimum utility across all web applications; and
- a web application executor that places web applications on one or more of the plurality of heterogeneous server machines.
4. The system of claim 1 wherein the request router classifies requests into flows depending on at least one of target web applications and service class.
5. The system of claim 4, wherein the request router places classified requests into per-flow queues within the request router.
6. The system of claim 5, wherein the request router dispatches requests from the per-flow queues based on a weighted-fair scheduling discipline that observes a concurrency limit of the system.
7. The system of claim 6, wherein the flow controller adjusts at least one of weights of the weighted-fair scheduling discipline and the concurrency limit of the system in response to at least one of changing workload intensity and system configuration.
8. The system of claim 7, wherein the flow controller builds a model of the system that allows it to predict the performance of a flow for any choice of concurrency limit and weights.
9. The system of claim 1, wherein the flow controller comprises a utility function calculator in communication with the placement controller.
10. The system of claim 1, wherein the job scheduler manages dependencies among jobs and performs resource matching of the jobs with one or more of the plurality of heterogeneous server machines.
11. The system of claim 1, further comprising:
- a web application workload profiler that obtains profiles for web application requests in the form of an average number of CPU cycles consumed by requests of a given flow, and provides the profiles for web application requests to the flow controller and placement controller; and
- a job workload profiler that obtains profiles for jobs in the form of at least one of the number of CPU cycles required to complete the job, the number of threads used by the job, and the maximum CPU speed at which the job may progress, and provides the profiles for jobs to the job scheduler.
12. The system of claim 1, wherein the each of the plurality of heterogeneous server machines comprise a plurality of domains.
13. The system of claim 1, wherein a first domain of the plurality of domains comprises a node agent in communication with the placement controller and the job scheduler.
14. The system of claim 13, wherein a second domain of the plurality of domains comprises a machine agent in communication with the node agent that manages virtual machines inside a given heterogeneous server machine.
15. The system of claim 14, wherein the node agent provides job management functionality within the heterogeneous server machine through interaction with the machine agent.
16. The system of claim 14, wherein the machine agent is capable of at least one of creating and configures a virtual machine image for a new domain, copying files from the second domain to another domain, starting a process in another domain, and controlling the mapping of physical resources to virtual resources.
Filed: May 3, 2012
Publication Date: Aug 30, 2012
Applicant: International Business Machines Corporation (Armonk, NY)
Inventors: David Michael Chess (Mohagan Lake, NY), Malgorzata Steinder (Fort Lee, NJ), Ian Nicholas Whalley (Pawling, NY)
Application Number: 13/463,371
International Classification: G06F 9/46 (20060101);