Electrical computers and digital processing systems: processing – Processing architecture – Distributed processing system
Reexamination Certificate
2000-06-06
2004-02-10
Tsai, Henry W. H. (Department: 2183)
Electrical computers and digital processing systems: processing
Processing architecture
Distributed processing system
C712S219000, C712S235000
Reexamination Certificate
active
06691220
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Technical Field
The present invention relates in general to data processing systems and in particular to processing of barrier operations in multiprocessor data processing systems. Still more particularly, the present invention relates to a method and processor architecture for improving processing of instruction sequences by speculatively executing load instructions before the completion of a prior synchronization (sync) operation.
2. Description of the Related Art
The need for faster and more efficient processing of computer instructions has typically been at the forefront of development in processors and data processing systems. Traditional processors execute instructions in a sequential order, i.e., subsequent instructions were executed only after the execution of the previous instruction was completed. These traditional processors are referred to as firmly consistent processors. In firmly consistent processors, there is an implied barrier after each instruction although there is no barrier operation included in the code sequence. Thus, these processors do not support speculative, out-of-order execution of instructions.
Development of faster processors necessitated the creation of weakly consistent processor architectures, which permit some amounts of speculation (such as branch speculation) and out-of-order execution of instructions. To enable these types of execution, a processor assigns a series of instructions to a group when no dependencies exist between instructions within that group. Instructions within a group can be executed in parallel or out-of-order (i.e., later instructions executed before earlier instructions). However, due to data dependencies, particularly with load and store instructions, within instructions sequences, instructions in different groups must be executed in program order to obtain correct processing results.
In multiprocessor systems, the completion of operations within code sequences executing on a first processor may be dependent on operations on a second processor. To observe such dependencies and maintain proper instruction execution order, it is necessary to place barrier instructions within the instruction sequence, which ensure that all instructions within a first code segment are fully executed (i.e., visible to all other processors) before any instruction within a subsequent code segment is executed. Barrier instructions are particularly necessary when the multiprocessor system includes superscalar processors supporting out-of-order instruction execution and weak memory consistency. For example, with load and store instructions executed by the load/store unit (LSU), a previous instruction that stores a value to a particular location must be executed before a later instruction that loads the value of that location. The instructions set architecture (ISA) supported by most popular commercial processors includes an instruction for setting such a processing boundary (i.e., a barrier instruction), which initiates a barrier operation on the system. In the PowerPC™ family of processors, for example, the barrier instruction, which is employed by a programmer to establish a processing boundary, is called a “sync” instruction, and the corresponding transaction on the system bus is called a synchronization operation (sync op). The sync instruction orders instruction execution. All instructions initiated prior to the sync instruction must be completed before the sync operation yields a sync acknowledgment (ack), which is returned to the LSU. In addition, no subsequent instructions are issued until the sync ack is received by the LSU. Thus, the sync instruction creates a boundary having two significant effects: (1) instructions which follow the sync instruction within the instruction stream will not be executed until all instructions which precede the sync instruction in the instruction stream have completed; and (2) instructions following a sync instruction within the instruction stream will not be reordered for out-of-order execution with instructions preceding the sync instruction.
An example instruction sequence with a sync op is illustrated in FIG.
6
A. The first three load or store (Ld/St) instructions before the sync instruction comprise group A, and the instructions following the sync instruction comprise group B. During normal operation, the above instruction sequence is executed serially, i.e., group A instructions are issued, followed by the sync. Upon detection of the sync, a sync operation is initiated on the system bus. The sync op prevents instructions in all of the processors to be held up (i.e., not issued) until the sync op is completed. The LSU then waits for the receipt of a sync ack from the indicating completion of all group A operations, before allowing any of the Group B instructions to be issued. Once the sync ack is received, the instructions from Group B are issued.
Because the operation of one processor in a multiprocessor system may affect another processor cache (i.e., data being loaded by one processor may have been invalidated by a store operation from another processor), sync operations, by ensuring that all previous store operations are completed, guarantee the processor will load valid data. The sync operation creates a boundary to further load and store operations by the issuing processor until all previous operations are completed. In doing so, the sync clears all processor pipelines of instructions.
In slower processors, which operate at, for example, 100 MHz, each barrier instruction, such as a sync, requires approximately 10 processor cycles to complete. With faster processors, however, such as those operating in the Ghz range, a sync completes in approximately 100 processor cycles. Thus, syncs place a significant burden on processor efficiency, particularly because, in typical software, syncs occur every 500-1000 instructions. Each occurrence of a sync causes processors in a data processing system to stall for 100 cycles while the issuing processor waits on the sync operation to complete. Another factor influencing sync performance in current architectures is the fact that the processor pipeline may include as many as 20-50 pipeline stages that require as many as 40-50 cycles to process an instruction. Thus, waiting until the entire pipeline drains (i.e., all previous instructions are completed) when a sync is encountered significantly compromises processor speed and efficiency.
The penalty associated with each sync operation worsens as technology progresses because of the rigid functional placement of sync operations. Other technological advances, such as, for example, increasing the number of execution units within a processor to allow more instructions to be executed in parallel or implementation of larger caches, resulting in more cache hits, positively affect processor efficiency. However, even if sync instructions remain a fixed percentage of all runtime instructions, because more instructions are being executed in parallel, the sync instructions consume a larger portion of available processor cycles and bandwidth. Furthermore, as memory hierarchies—all levels of which are affected by a sync instruction—become deeper, the performance penalty associated with a single sync instruction also increases.
The present invention recognizes that it would therefore be desirable to provide a method and processor architecture for enabling speculative execution of load instructions beyond a sync to reduce processor stalls while awaiting a sync ack and thereby increase processor speed and efficiency.
SUMMARY OF THE INVENTION
A processor architecture is described, which comprises a load store unit (LSU) coupled to an upper and a lower level cache and to processor registers. The processor architecture enables the processor to efficiently process load/store instruction sequences without losing processor cycles while waiting on the completion of a barrier operation (sync op). The LSU contains a barrier operation controller, which is coupled to and interacts with the LSU's load requ
Arimilli Ravi Kumar
Dodson John Steven
Guthrie Guy Lynn
Williams Derek Edward
Bracewell & Patterson L.L.P.
International Business Machines - Corporation
Salys Casimer K.
Tsai Henry W. H.
LandOfFree
Multiprocessor speculation mechanism via a barrier... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Multiprocessor speculation mechanism via a barrier..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Multiprocessor speculation mechanism via a barrier... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3301588