摘要:
A hosting partition update mechanism allows updating the I/O capability of a logically-partitioned computer system in a way that minimally affects the performance and availability of I/O in the computer system. When an update is needed, a new hosting partition is created with the desired update(s). The I/O adapters in the current hosting partition are then migrated to the new hosting partition. The migration of an I/O adapter from the current hosting partition to the new hosting partition is relatively fast, thereby minimally impacting system performance and availability of I/O. Once all of the I/O adapters have been migrated to the new hosting partition, the current hosting partition may be kept as a backup, or may be eliminated. Providing a new or backup hosting partition allows updates to be performed in the new or backup hosting partition in a non-disruptive manner while the current hosting partition continues to service I/O requests.
摘要:
A hosting partition update mechanism allows updating I/O capability of a logically-partitioned computer system in a way that minimally affects the performance and availability of I/O in the computer system. When an update is needed, a new hosting partition is created with the desired update(s). I/O adapters in the current hosting partition are then migrated to the new hosting partition. The migration of an I/O adapter from the current hosting partition to the new hosting partition is relatively fast, thereby minimally impacting system performance and availability of I/O. Once all of the I/O adapters have been migrated to the new hosting partition, the current hosting partition may be kept as a backup, or may be eliminated. Providing a new or backup hosting partition allows updates to be performed in the new or backup hosting partition in a non-disruptive manner while the current hosting partition continues to service I/O requests.
摘要:
Methods and systems for discovering whether a given connection to an InfiniBand port is a standard InfiniBand connection or a non-standard connection (e.g., indicating the presence of a proprietary device) are provided. The discovery may be performed by an end node, such that the end node interoperates with all standard InfiniBand components. Specific actions that are vendor unique, and potentially not compliant with the InfiniBand architecture, may not be done until after the discovery is complete and it has been verified that the noncompliant action will only be directed to entities know to be capable of processing them. These actions may include assuming the configuration responsibilities that would have been performed by the Subnet Manager in a standard InfiniBand network
摘要:
An I/O adapter sharing mechanism in a logically partitioned computer system allows sharing a network I/O adapter between logical partitions in an efficient manner. A first logical partition owns (or controls) the I/O adapter, and a second logical partition desires to use (or share) the I/O adapter. An I/O adapter device driver that includes a hardware interface is provided in the first logical partition. A virtual device driver is provided in the second logical partition that provides a set of functions that is at least partially defined by querying the I/O adapter device driver in the first logical partition. The I/O adapter sharing mechanism includes a transfer mechanism that allows data to be transferred directly from the virtual device driver in the second logical partition to the I/O adapter without passing through the I/O adapter device driver in the first logical partition.
摘要:
A method, apparatus and computer program product are provided for implementing queue pair connection protection over an interconnect network, such as InfiniBand. A message packet is received for a queue pair (QP) and the QP is checked for an imminent connection failure. Responsive to identifying an imminent connection failure, a special message processing mode is established for the QP. After the special message processing mode is established, packets of the message are received without establishing a message queue entry and without storing packet data.
摘要:
An apparatus and method for an advanced multiplexing technique to allow a single host to support multiple Internet Protocol (IP) queue pairs with little or no overhead are provided. With the apparatus and method, after a queue pair is created, Internet Protocol filter attributes and values are set up for the queue pair through value added features to the standard InfiniBand “QP Modify” method. The IP filters are used during normal operations to determine which queue pair is associated with an incoming packet. During normal operations, when a channel adapter receives an Internet Protocol (IP) over InfiniBand (IB) packet, it uses one or more of several fields in the packet's transport and/or network header to determine which queue pair shall receive the packet. Thus, the host channel adapter uses the IP filters to route incoming packets to the appropriate queue pair and thereby allow more than one queue pair to be used to support IP.
摘要:
An interrupt and message batching apparatus and method reduces the number and frequency of processor interrupts and resulting context switches by grouping I/O completion events together with a single processor interrupt in a manner that balances I/O operation latency requirements with processor utilization requirements to optimize overall computer system performance. The invention sends a message from a processor complex to an I/O adapter on an I/O bus commanding an I/O device connected to the I/O adapter to perform a function. Upon completion of the commanded function, the message processor in the I/O adapter generates a message and sends it to the processor complex on the I/O bus. The message is enqueued in the message queue of the memory, a message count is updated, and processor complex interrupt is signalled if and when the message count exceeds a message pacing count. A signalling timer may also be programmed with a fast response time value if the message has a relatively high latency or with a slow response time value if the message has a relatively low latency. The signalling timer is started when the message is enqueued and the processor complex interrupt is then signalled when the message count exceeds the message pacing count or when the signalling timer has elapsed.