Abstract: A coherence transformer for allowing a computer node and one or more external devices to share memory blocks having local physical addresses at a memory module of the computer node. The coherence transformer includes logic for ascertaining whether a memory access request from the external device for a memory block should be responded to using a snoop-only approach or an Mtag-only approach. The snoop-only approach requires a tag in a snoop tag array of the coherence transformer be available to track the memory block for an entire duration that the memory block is cached by the external device. The Mtag-only approach only temporarily stores the memory block until a global state associated with the memory block can be written back into the memory module of the computer node. The snoop tag array allows the coherence transformer to snoop the bus of the computer node to intervene and respond to memory access requests pertaining to a memory block externally cached and tracked by the snoop tag array.

Abstract: Memory space in the lower-level cache (LLC) of a computer system is allocated in cache-line sized units, while memory space in the higher-level cache (HLC) of the computer system is allocated in page sized units; with each page including two or more cache lines. Accordingly, during the execution of a program, cache-line-sized components of a page-sized block of data are incrementally stored in the cache lines of the LLCs. Subsequently, the system determines that it is time to review the allocation of cache resources, i.e., between the LLC and the HLC. The review trigger may be external to the processor, e.g., a timer interrupting the processor on a periodic basis. Alternatively, the review trigger may be from the LLC or the HLC, e.g., when the LLC is full, or when usage of the HLC drops below a certain percentage.

Abstract: A technique for system memory space address mapping in a multiprocessor computer system is provided. The disclosed mapping architecture may be applied to a multiprocessor computer system having multiple processing nodes (SMP nodes), where each processing node may include multiple processors. The system memory address space is split into different regions such that each of the n SMP nodes is assigned 1/n of the total address space. Cache coherency state information is stored for the memory in each SMP node. Memory regions may further be assigned to operate in one of three modes: normal, migratory, or replicate. When operating in normal mode, transaction to an address space assigned to a particular node are tried only locally in that node first. Transactions may be sent globally to other nodes if an improper cache coherency state is returned or if the address corresponds to a memory region assigned to another node. In migratory mode transactions are always sent globally.

Abstract: An architecture for an extended multiprocessor (XMP) computer system is provided. The XMP computer system includes multiple SMP nodes. Each SMP node includes an XMP interface and a repeater structure coupled to the XMP interface. The SMP nodes are connected to each other by unidirectional point-to-point links. The repeater structure in each SMP node includes an upper level bus, one or more transaction repeaters coupled to the upper level bus. Each transaction repeater broadcasts transactions to bus devices attached to a lower level bus, wherein each transaction repeater is coupled to a separate lower level bus. Transaction repeater includes a queue and a bypass path. Transaction originating in a particular SMP node are stored in the queue, whereas transactions originating in other SMP nodes bypass the incoming queue to the bus device. Multiple transactions may be simultaneously broadcast across the point-to-point link connections between the SMP nodes.

Abstract: A computer system includes a directory at each node which stores coherency information for the coherency units for which that node is the home node. In addition, the directory stores a data access state corresponding to each coherency unit which indicates the data access pattern observed for that coherency unit. The data access state may indicate migratory or non-migratory data access patterns. If the coherency unit has been observed to have a migratory data access pattern, then read/write access rights are granted. Conversely, if the coherency unit has been observed to have non-migratory data access patterns, then read access rights are granted. The home node further detects the migratory and non-migratory data access patterns and selects transitions between the migratory and non-migratory data access states independent of the cache hierarchies within the nodes which access the affected coherency unit. In one embodiment, a pair of counters are employed for each coherency unit.