Patent Application
20250224991
This application relates to high-performance computing (HPC).
HPC environments include clusters of servers, potentially on the order of hundreds of servers, that are networked together to solve large analysis problems. Currently, users of HPC environments must decide the technical details of when and where their job runs. Also, users increasingly have more service provider choices for these runs. Previously, users could only choose to run their job on-premises in an HPC queue. Today, those users can run jobs on-premises in multiple queues or run it with multiple cloud providers with varying compute capacity. Moreover, the cost of those cloud providers can change by the minute and, if the user is willing to wait, the cost may fall in the near future, as prices are known to fluctuate based on demand. This presents challenges for users to select an optimal HPC environment for their HPC computing jobs.
A computer-implemented method for high-performance computing (HPC) according to an example embodiment of the present disclosure includes receiving, from a user, a computing job request that includes or describes input data required for performing a computing job, and includes an urgency request for the computing job. The method also includes determining, for a plurality of HPC environments, which includes an in-house HPC environment associated with the user and a plurality of third-party HPC environments, an extent to which the plurality of HPC environments can perform the computing job and fulfill the urgency request. The method also includes, based on the determining, presenting to the user a summary of a cost and availability of each of the plurality of HPC environments for performance of the computing job according to the urgency request; receiving a selection of one of the plurality of HPC environments from the user based on the summary; and allocating the computing job to the selected HPC environment.
In a further embodiment of the foregoing embodiment, the determining and presenting are performed based on the in-house HPC environment not having sufficient computing capacity available to complete the computing job and fulfill the urgency request. The method includes, based on the in-house HPC environment having sufficient computing capacity available to complete the computing job and fulfill the urgency request, automatically allocating the computing job to the in-house HPC environment.
In a further embodiment of any of the foregoing embodiments, the determining includes polling the plurality of HPC environments to determine pricing and availability of the plurality of HPC environments, and the summary includes a ranking of the plurality of HPC environments based on an extent to which the plurality of HPC environments can complete the computing job according to the urgency request and based on cost.
In a further embodiment of any of the foregoing embodiments, the method includes determining a set of computing resources required for performing the computing job, which includes receiving a description of the set of computing resources required for the computing job as part of the computing job request. The method also includes determining the summary based on the determined set of computer resources required.
In a further embodiment of any of the foregoing embodiments, the method also includes determining a set of computing resources required for performing computing job, which includes estimating the set of computing resources required for the computing job based on the computing job request. The method also includes determining the summary based on the estimated set of computer resources required.
In a further embodiment of any of the foregoing embodiments, the computing job request includes one or more of: a number of computing cores needed for the computing job, a type of computing core needed for the computing job, an amount of memory needed for the computing job, an estimated length of the computing job, an amount of data storage needed for the computing job, and an amount data transfer needed for uploading the input data to the HPC environment and for downloading output data of the computing job from the HPC environment.
In a further embodiment of any of the foregoing embodiments, the computing job request includes a data locality requirement indicating one or more geographic restrictions on transfer of data associated with the computing job. The method includes, based on the data locality requirement, excluding an HPC environment that is unable to comply with the data locality requirement from the summary.
In a further embodiment of any of the foregoing embodiments, the summary includes a plurality of configurations for at least one of the third-party HPC environments that vary in terms of estimated completion date.
In a further embodiment of any of the foregoing embodiments, the method includes utilizing a machine learning algorithm trained with historical data of computing jobs performed by one of the plurality of HPC environments to predict future availability of computing resources at the one of the plurality of HPC environments.
In a further embodiment of any of the foregoing embodiments, for each of at least one of the plurality of HPC environments that can perform the computing job but cannot also fulfill the urgency request, the summary includes a best effort option for the HPC environment that indicates an earliest time the computing job could be completed by the HPC environment.
A computing device according to an example embodiment of the present disclosure includes processing circuitry operatively connected to memory and configured to receive, from a user, a computing job request that includes or describes input data required for performing a computing job, and includes an urgency request for the computing job. The processing circuitry is also configured to determine, for a plurality of HPC environments, which includes an in-house HPC environment associated with the user and a plurality of third-party HPC environments, an extent to which the plurality of HPC environments can perform the computing job and fulfill the urgency request. The processing circuitry is also configured to, based on the determination, present to the user a summary of a cost and availability of each of the plurality of HPC environments for performance of the computing job according to the urgency request. The processing circuitry is also configured to receive a selection of one of the plurality of HPC environments from the user based on the summary and allocate the computing job to the selected HPC environment.
In a further embodiment of the foregoing embodiment, the processing circuitry is configured to perform the determination and the presentation of the summary based on the in-house HPC environment not having sufficient computing capacity available to complete the computing job and fulfill the urgency request. The processing circuitry is also configured to, based on the in-house HPC environment having sufficient computing capacity available to complete the computing job and fulfill the urgency request, automatically allocate the computing job to the in-house HPC environment.
In a further embodiment of any of the foregoing embodiments, to determine the extent to which the plurality of HPC environments can perform the computing job and fulfill the urgency request, the processing circuitry is configured to poll the plurality of HPC environments. The summary includes a ranking of the plurality of HPC environments based on an extent to which the plurality of HPC environments can complete the computing job according to the urgency request and based on cost.
In a further embodiment of any of the foregoing embodiments, the processing circuitry is configured to determine a set of computing resources required for performing the computing job, which includes receiving a description of the set of computing resources required for the computing job as part of the computing job request. The processing circuitry is also configured to determine the summary based on the determined set of computer resources required.
In a further embodiment of any of the foregoing embodiments, the processing circuitry is configured to determine a set of computing resources required for performing the computing job, which includes estimating the set of computing resources required for the computing job based on the computing job request. The processing circuitry is also configured to determine the summary based on the estimated set of computer resources required.
In a further embodiment of any of the foregoing embodiments, the computing job request includes one or more of a number of computing cores needed for the computing job, a type of computing core needed for the computing job, an amount of memory needed for the computing job, an estimated length of the computing job, an amount of data storage needed for the computing job, and an amount data transfer needed for uploading the input data to the HPC environment and for downloading output data of the computing job from the HPC environment.
In a further embodiment of any of the foregoing embodiments, the computing job request includes a data locality requirement indicating one or more geographic restrictions on transfer of data associated with the computing job. The processing circuitry is configured to, based on the data locality requirement, exclude an HPC environment that is unable to comply with the data locality requirement from the summary.
In a further embodiment of any of the foregoing embodiments, the summary includes a plurality of configurations for at least one of the third-party HPC environments that vary in terms of estimated completion date.
In a further embodiment of any of the foregoing embodiments, the processing circuitry is configured to utilize a machine learning algorithm trained with historical data of computing jobs performed by one of the plurality of HPC environments to predict future availability of computing resources at the one of the plurality of HPC environments.
In a further embodiment of any of the foregoing embodiments, for each of at least one of the plurality of HPC environments that can perform the computing job but cannot also fulfill the urgency request, the summary includes a best effort option for the HPC environment that indicates an earliest time the computing job could be completed by the HPC environment.
The embodiments, examples, and alternatives of the preceding paragraphs, the claims, or the following description and drawings, including any of their various aspects or respective individual features, may be taken independently or in any combination. Features described in connection with one embodiment are applicable to all embodiments, unless such features are incompatible.
The computing job request 16 includes or describes input data required for performing a computing job and describes what computing operations are being requested (e.g., the request 16 includes or describes what program(s) or script(s) should be executed on what input data set and/or indicates where those items can be download). The description may include, for example, a link to the input data and/or a description of a size of the input data. The input data may include one or more computer models to be processed, for example.
The computing job request 16 also includes an urgency request for the computing job that indicates a desired processing time for the computing job (e.g., as a requested completion date and/or requested completion time).
In one or more embodiments, the computing job request 16 also includes one or any combination of the following:
The orchestration server 12 determines a set of computing resources required for performing the computing job described in the request 16. In one or more embodiments, this is determined based on one or more of: a description provided in the computing job request 16, or the orchestration server 12 estimating the amount of computing resources required for performing the computing job based on the computing job request 16.
The orchestration server 12 determines, for a plurality of HPC environments, including the in-house HPC environment 18 and the plurality of third-party HPC environments 20A-N, an extent to which the plurality of HPC environments can perform the computing job request and fulfill the urgency request. This may include polling the plurality of HPS environments, for example. Based on the determining, the orchestration server 12 presents to the user 13 a summary 30 of the cost and availability of each of the plurality of HPC environments for performance of the computing job according to the urgency request. The user 13 selects one of the plurality of HPC environments, and the orchestration server 12 then allocates the computing job to the selected HPC environment.
In the example of
Assume also that the requested urgency is one day. This causes results 38A-B to be provided first, even though the cost of some other providers is lower per core hour (e.g., results 38E (price A2) and 38G (price E1)).
In one or more embodiments, the computing job request 16 indicates the relative importance of cost vs. urgency in the form of a weighting and/or the user 13 indicates whether they have flexibility in the urgency (e.g., 1 day preferred, but 2-5 days acceptable), and these factor(s) are taken into account by the orchestration server 12 when providing the summary 30. If cost is weighted higher than availability, for example, then a lower cost result that is slower may be ranked higher than a more expensive result that is faster (e.g., result 38G higher than result 38A).
The example summary 30 of
Assume in one example that the indicated urgency of the request 16 is two days. Provider C and Provider E can perform the requested computing job but cannot do so while fulfilling the urgency request. Therefore, for these providers, the orchestration server 12 provides results 38D and 38F as a “best effort” option that indicates an earliest time the computing job could be completed by those HPC environment providers.
The orchestration server 12 determines a set of computing resources required for performing the computing job (step 64). In one or more embodiments, the orchestration server 12 makes the determination by parsing the computing job request 16, which describes the resources required. In one or more further embodiments, the orchestration server 12 determines the set of computing resources required for performing the computing job by estimating the set of computing resources required for the computing job based on the nature computing job request (e.g., based on the nature of the computational operations being requested).
The orchestration server 12 determines, for a plurality of HPC environments (step 66), which includes the in-house HPC environment 18 associated with the user 13 and a plurality of third-party HPC environments 20A-N, an extent to which the plurality of HPC environments can perform the computing job and fulfill the urgency request. The determination may include polling each of the HPC environments, for example (which may be beneficial as prices can fluctuate frequently).
Based on the determining, the orchestration server 12 presents to the user 13 the summary 30 of a cost and availability of each of the plurality of HPC environments for performance of the computing job according to the urgency request (step 68).
The orchestration server 12 receives a selection of one of the plurality of HPC environments from the user 13 based on the summary (step 70), and allocates the computing job to the selected HPC environment (step 72).
In one or more embodiments, the determining of step 66 and presenting of step 68 are performed based on the in-house HPC environment 18 not having sufficient computing capacity available to complete the computing job and fulfill the urgency request, and the method 60 includes, based on the in-house HPC environment 18 having sufficient computing capacity available to complete the computing job and fulfill the urgency request, automatically allocating the computing job to the in-house HPC environment 18.
In one or more embodiments, the computing job request 16 includes a data locality requirement indicating one or more geographic restrictions on transfer of data associated with the computing job (e.g., do not export data outside of the United States), and the method 60 includes, based on the data locality requirement, excluding from the summary 30 an HPC environment that is unable to comply with the data locality requirement (e.g., because it would use servers outside the United States).
In one or more embodiments, the orchestration server 12 utilizes a machine learning algorithm (e.g., a neural network) trained with historical data of computing jobs performed by one of the plurality of HPC environments to predict future availability and/or pricing of computing resources at the one of the plurality of HPC environments. Referring again to
Use of the orchestration server 12 described herein may simplify the process of submitting HPC computing jobs for users, improve maintainability by pointing various engineering workflows to a single tool (the orchestration server 12), reduce cost to run analyses (by facilitating the best options by users), and may significantly reduce the likelihood of vendor lock-in.
Although example embodiments have been disclosed, a worker of ordinary skill in this art would recognize that certain modifications would come within the scope of this disclosure. For that reason, the following claims should be studied to determine the scope and content of this disclosure.
