Precomputation for data center load balancing
원문보기
IPC분류정보
국가/구분
United States(US) Patent
등록
국제특허분류(IPC7판)
G06F-009/46
G06F-009/48
출원번호
US-0492385
(2009-06-26)
등록번호
US-8839254
(2014-09-16)
발명자
/ 주소
Horvitz, Eric J.
Belady, Christian L.
출원인 / 주소
Microsoft Corporation
대리인 / 주소
Choi, Dan
인용정보
피인용 횟수 :
6인용 특허 :
35
초록▼
Pre-computing a portion of forecasted workloads may enable load-balancing of data center workload, which may ultimately reduce capital and operational costs associated with data centers. Computing tasks performed by the data centers may be analyzed to identify computing tasks that are eligible for p
Pre-computing a portion of forecasted workloads may enable load-balancing of data center workload, which may ultimately reduce capital and operational costs associated with data centers. Computing tasks performed by the data centers may be analyzed to identify computing tasks that are eligible for pre-computing, and may be performed prior to an actual data request from a user or entity. In some aspects, the pre-computing tasks may be performed during a low-volume workload period prior to a high-volume workload period to reduce peaks that typically occur in data center workloads that do not utilize pre-computation. Statistical modeling methods can be used to make predictions about the tasks that can be expected to maximally contribute to bottlenecks at data centers and to guide the speculative computing.
대표청구항▼
1. A method implemented by one or more processors of load-balancing a data center workload of a data center, the method comprising: predicting a high-volume period having a forecasted workload that exceeds a target workload level;ascertaining a low-volume period prior to the high-volume period based
1. A method implemented by one or more processors of load-balancing a data center workload of a data center, the method comprising: predicting a high-volume period having a forecasted workload that exceeds a target workload level;ascertaining a low-volume period prior to the high-volume period based at least in part on an inverse curve of a cumulative curve, the cumulative curve being an overlay of data center operational factors that indicate a price per operation unit for operating the data center;selecting computing tasks that are anticipated to occur during the forecasted workload for pre-computation;performing the pre-computation of the computing tasks associated with the forecasted workload during the low-volume period to generate manipulated data; storing the manipulated data in cache; andretrieving the manipulated data upon a request that uses the manipulated data during the high-volume period to expedite processing of the request. 2. The method as recited in claim 1, wherein the data center operational factors indicate efficient periods to schedule workloads for pre-computation that are preferentially selected over non-efficient periods. 3. The method as recited in claim 1, wherein the data center operational factors further include a cost of power from a power grid that is shared by the data center and another power consumer and another cost of power from a dedicated power supplier that exclusively supplies power to the data center. 4. The method as recited in claim 1, wherein the pre-computation includes at least one of: pre-computing anticipated queries;pre-computing anticipated transactions; andoptimizing databases to enable expedited results of database queries. 5. The method as recited in claim 1, wherein the selecting computing tasks is based on a historical analysis of tasks that are anticipated to reoccur on one or more future dates. 6. The method as recited in claim 1, wherein the selecting occurs at near-real time as the data center workload enters the high-volume period. 7. A data center processing system, comprising: a communication adapter configured to enable a data center to receive a request from an entity computing device and transmit a response to the entity computing device;one or more processors; andmemory storing computer readable instructions executable by the one or more processors, wherein the computer readable instructions provide a decomposer module configured to:identify computing tasks to be pre-computed in anticipation of the request expected to occur during a high-volume period of workload of the data center, the computing tasks to fulfill at least a portion of the request;ascertain a low-volume period based at least in part on an inverse curve of a cumulative curve, the cumulative curve being an overlay of data center operational factors that indicate a price per operation unit for operating the data center;pre-compute the computing tasks during the low-volume period to create manipulated data based on a computing blueprint that uses historical computing trends to determine how to pre-compute the computing tasks;store the manipulated data in cache for access during the high-volume period; andretrieve a portion of the manipulated data to expedite satisfying the request upon receipt of the request. 8. The system as recited in claim 7, wherein the decomposer module is further configured to delete unused instances of the manipulated data when the manipulated data becomes stale. 9. The system as recited in claim 7, wherein the computing tasks include queries that are pre-computed by selecting keywords from an early data release. 10. The system as recited in claim 7, wherein the computing tasks include transactions that are pre-computed by inputting at least one of user information, item information, and payment information. 11. The system as recited in claim 7, wherein the computing tasks include an optimization of databases to enable expedited results of database queries. 12. The system as recited in claim 7, wherein the instructions further provide a scheduler module configured to: predict workload volume of the data center; andschedule the pre-compute to occur during the low-volume period. 13. The system as recited in claim 12, wherein the scheduler module is further configured to provide the overlay of the data center operational factors that affect workload scheduling. 14. A method implemented by one or more processors of processing a portion of a computing workload of a data center prior to a request of the computing workload, the method comprising: identifying computing tasks performed by the data center to be pre-computed prior to an initiated request;obtaining inputs for pre-computing based on a historically based blueprint that uses historical trends to predict future computing demands;ascertaining a low-volume workload period of the data center based at least in part on an inverse curve of a cumulative curve, the cumulative curve being an overlay of data center operational factors that indicate a price per operation unit for operating the data center, the inverse curve having at least one high curve portion that is equivalent to a mean workload of the data center;performing the pre-computing of computing tasks during the low-volume workload period of the data center prior to a high-volume workload period, the pre-computing tasks to create manipulated data;storing the manipulated data in cache; andretrieving the manipulated data in response to a received request that corresponds to the computing tasks. 15. The method as recited in claim 14, wherein the inputs include keywords selected from an early data release that are search terms for queries. 16. The method as recited in claim 14, wherein the inputs include transaction information to enable pre-computing of a transaction by performing at least one of compiling a portion of a transaction, adding items to an order, processing a payment, or assigning a customer. 17. The method as recited in claim 14, further comprising generating a forecasted schedule of the computing workload of the data center that includes the low-volume workload period and the high-volume workload period. 18. The method as recited in claim 17, wherein the operational factors include at least one of a cooling cost for one or more servers in the data center, a cost of power from a power grid that is shared by the data center and another power consumer, or another cost of power from a dedicated power supplier that exclusively supplies power to the data center. 19. The method as recited in claim 14, wherein the identifying includes identifying the computing tasks performed by the data center based on statistical models using at least one of a user history, machine usage, or data center performance.
연구과제 타임라인
LOADING...
LOADING...
LOADING...
LOADING...
LOADING...
이 특허에 인용된 특허 (35)
Graybill, David P.; Iyengar, Madhusudan K; Newcomer, Jeffrey A.; Schmidt, Roger R., Airflow recirculation and cooling apparatus and method for an electronics rack.
Davis Glenn A. (Lilburn GA) Massara James M. (Dunwoody GA), Apparatus and method for controlling an electrical load and monitoring control operations and the electrical load.
Karaoguz, Jeyhan; Currivan, Bruce; Diab, Wael William; Kim, Yongbum; Ma, Kenneth; Johas Teener, Michael, Method and system for power-limited switching and/or routing in a network.
Kuelbs,Gustav P.; Kuelbs,John A.; Kuelbs,Gregory G.; Domzalski,Donald, Method of producing, selling, and distributing articles of manufacture through the automated aggregation of orders and the visual representation of standardized shipping volumes.
Ritter, Allen Michael; Harbourt, Cyrus David; Wagoner, Robert Gregory, Methods for coupling an energy storage system to a variable energy supply system.
Belady, Christian L.; Larus, James R.; Reed, Danny A.; Borgs, Christian H.; Chayes, Jennifer Tour; Lobel, Ilan; Menache, Ishai; Nazerzadeh, Hamid; Jain, Navendu, Data center system that accommodates episodic computation.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.