Sun Microsystems, Inc.

Marshall C Pease

Madhusudhan Talluri

Srinivasan Viswanathan

When a client computer requests data from a disk or similar device at a server computer, the client exports the memory associated with an allocated read buffer by generating and storing one or more incoming MMU (IMMU) entries that map the read buffer to an assigned global address range. The remote data read request, along with the assigned global address range is communicated to the server node. At the server, the request is serviced by performing a memory import operation, in which one or more outgoing MMU (OMMU) entries are generated and stored for mapping the global address range specified in the read request to a corresponding range of local physical addresses. The mapped local physical addresses in the server are not locations in the server's memory. The server then performs a DMA operation for directly transferring the data specified in the request message from the disk to the mapped local physical addresses. The DMA operation transmits the specified data to the server's network interface, at which the mapped local physical addresses to which the data is transferred are converted into the corresponding global addresses. The specified data with the corresponding global addresses are then transmitted to the client node. The client converts the global addresses in the received specified data into the local physical addresses corresponding to the allocated receive buffer, and stores the received specified data in the allocated receive buffer.

System and method for efficient remote disk I/O

A first computer sends a sequence of messages to a second computer using remote write operations to directly store each message in a corresponding memory location in the second computer, without performing remote read operations to confirm storage of each message in the second computer's memory. The first computer detects message transmission errors, if any, during the transmission of each message to the second computer and when a message transmission error is detected, resends the message to the second computer. The first computer also sends trigger messages to the second computer to prompt the second computer to process messages stored in the second computer's memory. The second computer processes each received message and uses a remote write operation to store an acknowledgment message in a corresponding memory location in the first computer. The first computer, upon detecting a failure to receive the acknowledgment message corresponding to any of the previously sent messages, undertakes remedial actions to determine whether the second computer has processed each of the unacknowledged previously sent messages. When the remedial actions determine that the second computer has not processed the unacknowledged previously sent message, the first computer prompts the second computer to process the unacknowledged previously sent message. When the remedial actions determine that the second computer has already processed the unacknowledged previously sent message, the first computer stores the acknowledgment message in a corresponding location in the first computer's memory.

System and method for message transmission between network nodes using

In a distributed computer system, a sending node prompts a receiving node to allocate and export to the sending node one or more memory 'segments'. Each allocated segment is sufficiently large to hold multiple receive buffers whose size fall within a predefined range of receive buffer sizes. Once a segment has been allocated and exported, the sending node allocates receive buffers within the segment, using sequentially contiguous portions for successive receive buffers, without any interaction with the receiving node. Messages are transmitted to the receiving node by remotely writing the data portion of each message to an allocated receive buffer and writing a control message with a pointer to the corresponding receive buffer to a message queue in the receiving node. The receiving node processes messages within the portions of the allocated segments specified by each control message and does not keep track of the used and/or unused portions of each segment. Only the sending node keeps track of the unused portions of each allocated segment. As a result, the computational and communication resource overhead associated with allocating receive buffers is substantially reduced. Once the sending node has exhausted a segment, it sends either a segment release message or segment recycle message to the receiving node. Recycling the segment causes the receiving node to re-enable the sending node to use the segment once the receiving node has processed all the messages previously written to receive buffers in the segment.

Growing community of inventors

Mountain View, CA, United States of America

Marshall C Pease