Patent Application
20130061220
The following co-pending applications are incorporated herein by reference in their entireties: U.S. patent application Ser. Nos. 12/760,920, filed Apr. 15, 2010 and 12/876,623, filed Sep. 7, 2010.
The presently disclosed embodiments are directed to providing a system and method for the provisioning of computational resources in a computer cloud, and, more particularly, to providing a system and method of efficiently handling transient bursts of demand for computer processing resources.
Cloud computing is known in the art as becoming increasingly popular for efficiently distributing computer resources to entities which would otherwise not have access to large amounts of processing power. Among many other things, for example, cloud computing has been utilized as one solution to accommodate situations in which an entity (e.g., a company, individual, etc.) connected to a cloud of inter-connected computers requests a task or job to be completed that requires a large amount of computational processing power and that must be completed in a relatively short amount of time (a “burst” job). However, even with the generally more efficient resource management provided by cloud computing, it remains difficult to predict, maintain and/or provision for such burst demands of computational power.
For example, a sudden need of computational power for some kinds of real-time analysis, such as constructing a document redundancy graph or performing cross document paragraph matching, results in a burst of computational activity requiring a corresponding burst of computational processing power. That is, these burst jobs may be of short duration but require an extremely heavy processing load. “Time-sharing” of the resources of a virtual machine (VM) in a cloud is sometimes used, in which idle VMs associated with a first user or entity are used for processing burst loads from a second user or entity. Time-sharing between VMs is often impractical, however, due to security concerns, namely, the potential sharing of confidential information between two parties in the same cloud (e.g., the data from a first party cannot be processed using the processor of a second party because there is a risk that the second party could access, save, or copy the first party's data from the processor, temporary or log files associated with the processor, etc.). While on-demand scaling of the parameters of a virtual machine (e.g., memory, processing power, storage, etc.) in a cloud is typically used to provide flexibility to users in the cloud, it can take up to a few minutes (say, 3-4 minutes) to spin up or activate a new virtual machine to provide additional resources. Thus, scaling on-demand by spinning up additional VMs is impractical for burst computations which must be completed in less than a few minutes (i.e., <3-4 minutes) in order to meet quality of service (QoS) requirements, because the demand will disappear before a virtual machine has a chance to even start up. Some entities may handle burst processing needs by using an always-on approach, in which all virtual machines and all possible computational resources are always available. However, for many entities, bursts occur rarely enough (e.g., once an hour, day, week, etc.) to make the always-on approach inefficient and costly except for entities or organizations having large data centers (e.g., Google, Amazon, Microsoft, etc.). Thus, there is need for computational techniques that provision computing resources for burst loads in a cost-effective and secure manner while respecting QoS requirements.
Broadly, the methods discussed infra provide provisioning of computational or processing power for processing burst requests within a cloud. Burst requests are common, for example, in real-time document analysis. Intense document processing demands or other burst tasks can last for merely a minute on a large set of parallel computing resources. If the analysis is performed on a small set of computing resources, it often renders the results tardy enough to be useless. If there are a large number of such demands, additional resources need to be provisioned to keep response times down for both burst and typical jobs. However, additional resources cannot be provisioned after arrival of the burst request because it would take too much time to ready any additional resources. Additional resources cannot be provisioned based on predictions of load because burst tasks are inherently intermittent and transient. Such load has to be managed by using inter-cloud workload provisioning (over a number of clouds) and/or live migration of running workload. These methods are most applicable to companies that are averse to using permanent, always-on, or dedicated data centers for providing analysis workload.
According to aspects illustrated herein, there is provided a method for provisioning computing resources for handling bursts of computing power including creating at least one auxiliary virtual machine in a first cloud of a first plurality of interconnected computing devices having at least one processor, suspending the at least one auxiliary virtual machine, receiving a burst job requiring processing in a queue associated with at least one active virtual machine, transferring a workload associated with the queue from the at least one active virtual machine to the at least one auxiliary virtual machine, resuming the at least one auxiliary virtual machine, and processing the workload with the at least one auxiliary virtual machine.
According to other aspects illustrated herein, there is provided a method for provisioning computing resources to accommodate for bursts of processing power demand for at least one virtual machine including providing at least one virtual machine in a cloud wherein at least one processor is associated with the at least one virtual machine, and the at least one processor alternates between a busy phase and an idle phase while processing a first job, determining when the at least one processor is in the idle phase, receiving a burst job requiring a larger amount of processing by the at least one processor per unit time in comparison with the first job, and processing at least a portion of the burst job during at least one of the idle phases of the at least one processor.
Other objects, features and advantages of one or more embodiments will be readily appreciable from the following detailed description and from the accompanying drawings and claims.
Various embodiments are disclosed, by way of example only, with reference to the accompanying drawings in which corresponding reference symbols indicate corresponding parts, in which:
At the outset, it should be appreciated that like drawing numbers on different drawing views identify identical, or functionally similar, structural elements of the embodiments set forth herein. Furthermore, it is understood that these embodiments are not limited to the particular methodology, materials and modifications described and as such may, of course, vary. It is also understood that the terminology used herein is for the purpose of describing particular aspects only, and is not intended to limit the scope of the disclosed embodiments, which are limited only by the appended claims.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood to one of ordinary skill in the art to which these embodiments belong. As used herein, “cloud computing” is intended generally to mean the sharing of computer resources (e.g., memory, processing power, storage space, software, etc.) across several computers, machines, or servers through a network, such as the Internet. An evolving definition of cloud computing is provided by the National Institute of Standards and Technology. Thus, a “cloud” is generally meant to be a collection of such interconnected machines, computers, or computing devices. By “computer,” “PC,” or “computing device” it is generally meant any analog or digital electronic device which includes a processor, memory, and/or a storage medium for operating or executing software or computer code. By “virtual” or “virtual machine” it is meant a representation of a physical computer, such as having an operating system (OS) accessible by a user and usable by the user as if it were a physical machine, wherein the computing resources (e.g., memory, processing power, software, storage, etc.) are obtained from a shared pool of resources, such that each virtual machine may be a collection of resources from various inter-connected computing devices, or alternatively, may be a sub-set of resources from a single computing device. Virtual machines may be accessible internally or externally. As used herein, “external” or an “external cloud” is intended to mean at least one computer arranged in a different location than the entity requesting the computation to be performed. As used herein, “internal” or an “internal cloud” is intended to mean at least one computer arranged in the same location as the entity requesting the computing to be performed, including a plurality of computers interconnected to each other and the entity requesting the computation to be performed. As used herein, “inter-cloud” or “intercloud” may generally be used as an adjective to refer to having the quality or nature of multiple connected clouds or being communicated or transferred between clouds; or, as a noun to refer to an inter-connected cloud (or group) of clouds. A “burst” or “burst job” as used herein particularly refers to a computer job, task, request, or query that requires a relatively large amount of processing power that must be completed in a short amount of time. Alternatively stated, a burst job required a relatively large amount of processing per unit time for completion of the burst job in comparison to typical jobs. “Burst” also more generally reflects the nature of any job which a computer, cloud, or processor is unable to handle alone, and that must be “bursted” out to other computers, VMs, or processors in an internal cloud, or out to an external cloud, to meet quality of service requirements.
Moreover, although any methods, devices or materials similar or equivalent to those described herein can be used in the practice or testing of these embodiments, some embodiments of methods, devices, and materials are now described.
Referring now to the Figures,
Queues 16a-16c are also shown schematically in
In many scenarios, it is unacceptable that extra VMs are started only to wait permanently for additional overflow queries (e.g., an always-on scheme), as this is wasteful of resources.
As shown in
As jobs are requested to be performed on the active VMs, it is determined if the estimated wait time until completion for each queued job is greater than a QoS limit (step 28). The estimated wait time could be calculated or determined according to any means known in the art. For example, a control unit could take a summation of the size of all data that needs to be processed for completion of each job, and divide that sum by a known or empirically calculated average rate by which the active machines can process data per unit time in order to determine how long it will take the active VMs to complete each job in the queue. Regardless of how the estimated wait time is determined, the jobs are processed according to whichever job is the most time sensitive (step 30). By most time sensitive it is meant, for example, the job which is most at risk of not being timely completed. Alternatively or in addition, step 30 may include sending newly received tasks to the queue that is determined to have the shortest overall estimated wait time in step 28, so that heavily loaded queues do not get overloaded while unloaded queues remain empty. It should be appreciated that some other method or measure could be used to determine which job should be processed next in step 30 (e.g., “first-in, first-out”, “last-in, first-out”, smallest data size, other QoS measures, etc.). To ensure QoS requirements are not missed, step 28 could be run continuously or repeatedly.
If the wait time is determined to be impermissibly long, particularly as a result of a requested burst job, then the workload in the corresponding queue is suspended, and the corresponding virtual machines are written to disk (step 32). Simultaneously, the workload is transferred to a cloud and/or a control unit in a cloud, which is associated with a sufficient number of suspended auxiliary virtual machines to timely handle the task (step 34). The transferred workload may include all of the jobs in the queue at the time the active VMs are suspended in step 32, or merely the burst job, which takes priority over the other jobs. Part of the transferred workload is the additional computing that is required to perform the transfer. It is assumed that data communication in the cloud is sufficient for large data transfers, such as over a 10 Gigabit Ethernet system. Further, it is assumed that the jobs are processing intensive (such as many small html files) and not data intensive (such as processing large numbers of high-resolution tiff or other images).
The auxiliary VMs are then resumed from their suspended state and reinitialized and/or resynchronized with the other machines in the cloud (step 36). While steps 32, 34, and 36 are shown occurring sequentially, but it should be appreciated that these steps could occur simultaneously or in any other order. Typically, VMs can be resumed from a suspended state, resynchronized in the cloud, and ready for parallel computing within about ten seconds because all of the startup scripts have already been run to establish connections between the VMs in step 24. After the suspended VMs are resumed, the transferred workload is processed in parallel using the newly resumed VMs (step 38).
Next, it is checked to see whether there is a sustained load on the auxiliary VMs (step 40). If there is a sustained load (e.g., additional queries and/or jobs that must be processed immediately), then the original VMs can be resumed (step 42), and utilized for processing the transferred workload, the additional queries, the original jobs if only the burst job was transferred to the auxiliary VMs, etc. Alternatively, the original VMs could be restarted on a different set of machines for a different purpose entirely, such as to handle a new burst job from a different set of machines, while the auxiliary machines process all additional workload. If there is not a sustained load, then any unneeded auxiliary machines can be re-suspended, until they are needed at a later time (step 44). In this way, in some embodiments the active VMs can become auxiliary VMs that are suspended and waiting for typical or burst jobs that need processing, while auxiliary VMs can become active VMs for handling the processing of common non-burst tasks within the cloud. In other embodiments the auxiliary VMs are only used for processing burst loads on an as-needed basis and are re-suspended after processing each burst job.
For example, one scenario is now described with respect to
Thus, in this scenario, it could be determined in step 28 by control units 18a and 18c that the estimated wait time for jobs 20a-20d in queue 16a and jobs 20g-20j in queue 16c are satisfactory to meet QoS requirements, and therefore the jobs are be processed according to whichever job is the most time sensitive (or by whichever other metric or method is desired) in step 30. In this scenario, it could be also be determined by control unit 18b in step 28 that component 12b will not be able to timely complete job 20k if processed by only machines 14d and 14e, due to the large processing requirements.
Accordingly, in this example, control unit 18b would proceed to step 32. For component 12b, virtual machines 14d and 14e would be suspended in step 32 and the workload transferred in step 34 to component 12c, where there are more available suspended VMs, specifically, machines 14g, 14h, and 14i. This workload could include every job in queue 16b at the time (i.e., jobs 20e, 20f, and 20k) or just the burst job (i.e., job 20k) These resumed machines would then process the transferred workload in step 38. Control unit 18c would monitor queue 16c to see if there is sustained load for resumed VMs 14g, 14h, and 14i in step 40. If there is no sustained load, then any workload could be resumed on machines 14d and 14e, which are resumed in step 42, while the auxiliary machines are suspended in step 44 when they are no longer needed.
Alternatively, just a portion of the resumed auxiliary machines could be re-suspended, for example, one of machines 14g, 14h, or 14i, and the two remaining machines used to process the jobs that would have otherwise have been processed by machines 14d and 14e. As previously mentioned, the now-suspended machines which were originally active (e.g., machines 14d and 14e) could be used as auxiliary machines to handle further burst loads. For example, if it is determined by control unit 18c that machine 14f will not be able to process all of its jobs timely, the workload could be transferred to component 12b, and machines 14d and 14e resumed to process the transferred load. In this way, the VMs can be suspended, written to disk, and resumed as needed to handle burst loads from any point in cloud 10. As described previously, it should be appreciated that cloud 10 could represent a single cloud that bursts internally between computers or virtual machines within a cloud; or, cloud 10 could represent a cloud of clouds, where the jobs are bursted externally between clouds.
As discussed previously, in cloud computing arrangements, a plurality of virtual machines are established and operatively connected together to process massive amounts of data in parallel. If security concerns are not an issue, then the same virtual machines can be used to process jobs for more than one party, such that a burst job can be interleaved among a main task or plurality of smaller common tasks. That is, typically, the virtual machines are processing a main task or job, but may occasionally receive a burst job that requires a relatively large amount of processing to be completed within a short amount of time, such as discussed previously with respect to
Accordingly, a method of burst-interleaving is proposed with respect to the schematic illustration of
Thus, if a certain set of virtual machines across an entire cloud can be identified that have a similar utilization profile (e.g., they all generally follow profile 46), then a supplemental burst load can be interleaved between the busy phases. That is, as shown schematically in
Furthermore, burst interleaving could be similarly performed by pairing certain kinds of workload together. For example, if the main job is a “burst accommodative” workload, such as a batch processing job comprising a plurality of smaller jobs that is going to take several hours to perform, then idle phases could be “artificially” inserted after completion of each smaller job. That is, the processors or a controller for the processors could temporarily pause processing of subsequent jobs in the batch to check for a burst job, and if one is found, to process the burst job before the remainder of the batch. These artificial pauses could be inserted every specified number of completed jobs, once per given unit of time, etc. Thus, a batch job can be artificially structured to resemble profile 46, wherein the processing of each job within the batch is represented by crests 50, while the artificially implanted pauses between jobs are represented by troughs 52. Likewise, the processing of the supplemental or burst job would resemble profile 48. Because of the predictability of competition of tasks within the batch, there can be a high level of confidence that pauses can be inserted sufficiently often to ensure that bursts are handled timely and/or to meet any other QoS requirements. Further, since the method is intended only to handle infrequent burst demands, these bursts will not unduly delay the results of the batch process. Further, it is assumed that the workloads are being processed over a large set of nodes in a cloud, so any burst job can be handled quickly by the associated processors before returning to work on the batch process.
It will be appreciated that various aspects of the above-disclosed embodiments and other features and functions, or alternatives thereof, may be desirably combined into many other different systems or applications. Various presently unforeseen or unanticipated alternatives, modifications, variations or improvements therein may be subsequently made by those skilled in the art which are also intended to be encompassed by the following claims.
