摘要:
Methods and apparatus for reducing memory latency in a software application are disclosed. A disclosed system uses one or more helper threads to prefetch variables for a main thread to reduce performance bottlenecks due to memory latency and/or a cache miss. A performance analysis tool is used to profile the software application's resource usage and identifies areas in the software application experiencing performance bottlenecks. Compiler-runtime instructions are generated into the software application to create and manage the helper thread. The helper thread prefetches data in the identified areas of the software application experiencing performance bottlenecks. A counting mechanism is inserted into the helper thread and a counting mechanism is inserted into the main thread to coordinate the execution of the helper thread with the main thread and to help ensure the prefetched data is not removed from the cache before the main thread is able to take advantage of the prefetched data.
摘要:
Methods and apparatus for reducing memory latency in a software application are disclosed. A disclosed system uses one or more helper threads to prefetch variables for a main thread to reduce performance bottlenecks due to memory latency and/or a cache miss. A performance analysis tool is used to profile the software application's resource usage and identifies areas in the software application experiencing performance bottlenecks. Compiler-runtime instructions are generated into the software application to create and manage the helper thread. The helper thread prefetches data in the identified areas of the software application experiencing performance bottlenecks. A counting mechanism is inserted into the helper thread and a counting mechanism is inserted into the main thread to coordinate the execution of the helper thread with the main thread and to help ensure the prefetched data is not removed from the cache before the main thread is able to take advantage of the prefetched data.
摘要:
Method, apparatus and system embodiments to schedule user-level OS-independent “shreds” without intervention of an operating system. For at least one embodiment, the shred is scheduled for execution by a scheduler routine rather than the operating system. The scheduler routine may receive compiler-generated hints from a compiler. The compiler hints may be generated by the compiler without user-provided pragmas, and may be passed to the scheduler routine via an API-like interface. The interface may include a scheduling hint data structure that is maintained by the compiler. Other embodiments are also described and claimed.
摘要:
A method for improving parallel processing of computer programs. DOACROSS loops and similar code are identified and parallelized using a post-wait control structure. The post-wait control structure may be implemented to include any one of a single counter to enforce an order of execution, an array to track code completion that is indexed by a modulus of a positive integer number, and/or a set of arrays to track a last code completed by a thread and a current code being executed by a thread.
摘要:
A method for improving parallel processing of computer programs. DOACROSS loops and similar code are identified and parallelized using a post-wait control structure. The post-wait control structure may be implemented to include any one of a single counter to enforce an order of execution, an array to track code completion that is indexed by a modulus of a positive integer number, and/or a set of arrays to track a last code completed by a thread and a current code being executed by a thread.
摘要:
Method, apparatus and system embodiments to schedule user-level OS-independent “shreds” without intervention of an operating system. For at least one embodiment, the shred is scheduled for execution by a scheduler routine rather than the operating system. The scheduler routine may receive compiler-generated hints from a compiler. The compiler hints may be generated by the compiler without user-provided pragmas, and may be passed to the scheduler routine via an API-like interface. The interface may include a scheduling hint data structure that is maintained by the compiler. Other embodiments are also described and claimed.
摘要:
Disclosed are embodiments of a system, methods and mechanism for management and translation of mapping between logical sequencer addresses and physical or logical sequencers in a multi-sequencer multithreading system. A mapping manager may manage assignment and mapping of logical sequencer addresses or pages to actual sequencers or frames of the system. Rationing logic associated with the mapping manager may take into account sequencer attributes when such mapping is performed Relocation logic associated with the mapping manager may manage spill and fill of context information to/from a backing store when re-mapping actual sequencers. Sequencers may be allocated singly, or may be allocated as part of partitioned blocks. The mapping manager may also include translation logic that provides an identifier for the mapped sequencer each time a logical sequencer address is used in a user program. Other embodiments are also described and claimed.
摘要:
A first execution time of a first thread executing on a first processing unit of a multiprocessor is determined. A second execution time of a second thread executing on a second processing unit of the multiprocessor is determined, the first and second threads executing in parallel. Power is set to the first and second processing units to effectuate the first and second threads to finish executing at approximately the same time in future executions of the first and second threads. Other embodiments are also described and claimed.
摘要:
Data structure creation, organization and management techniques for data local to user-level threads are provided. Other embodiments are also described and claimed.
摘要:
A first execution time of a first thread executing on a first processing unit of a multiprocessor is determined. A second execution time of a second thread executing on a second processing unit of the multiprocessor is determined, the first and second threads executing in parallel. Power is set to the first and second processing units to effectuate the first and second threads to finish executing at approximately the same time in future executions of the first and second threads. Other embodiments are also described and claimed.