Method and apparatus for managing application state in a network interface controller in a high performance computing system

A method at a network node, the method comprising:

• with a network interface controller (NIC) of the network node, exchanging communications between the network node and a router of a network including the network node, the NIC coupled to a host of the network node;
• caching all levels of a page table hierarchy in the NIC, each level cached in an independent cache;
• and
• upon receiving a request including a virtual address and a process identifier, concurrently querying all of the independent caches.

A method at a network node, the method comprising:

• with a network interface controller (NIC) of the network node, exchanging communications between the network node and a router of a network including the network node, the NIC coupled to a host of the network node;
• monitoring resource exhaustion, by a predictor stage at the front of a processing pipeline of the NIC, the processing pipeline processing messages received from a network layer for delivery to an application;
• providing, by an event posting stage at the back of the processing pipeline, a prediction of when resource exhaustion is near, to the predictor stage;
• and
• stalling, by the predictor stage, processing of messages in the processing pipeline in response to the indication of resource exhaustion is near.

A method at a network node, the method comprising:

• with a network interface controller (NIC) of the network node, exchanging communications between the network node and a router of a network including the network node, the NIC coupled to a host of the network node;
• providing an event queue shared by the host and the NIC, the event queue to manage resources shared by the host and the NIC;
• allocating, by the host, receive queue entries in memory in the host for use by the NIC;
• sending, by the host, handles to the allocated receive queue entries in an event on the event queue;
• and
• freeing, by the NIC, the receive queue entries allocated by the host.

A method at a network node, the method comprising:

• with a network interface controller (NIC) of the network node, exchanging communications between the network node and a router of a network including the network node;
• and
• storing, in a retransmission buffer in the NIC, handles to host memory structures used by the NIC to retransmit a message over a communications network;
• overwriting one of the stored handles in the retransmission buffer with a handle to an entry of the retransmission buffer to convert the entry in the retransmission buffer to a network header, the network header used to retransmit a message over the communications network;
• and
• using the handle to the retransmission buffer to search for a retransmission buffer corresponding to a received network message acknowledgement.

A method at a network node, the method comprising:

• with a network interface controller (NIC) of the network node, exchanging communications between the network node and a router of a network including the network node, the NIC coupled to a host of the network node;
• tracking messages with the NIC, including:
• upon receiving a request from an initiator to transmit a first message to a target, adding a tracker entry associated with the initiator in a transmit message queue, the tracker entry to store a sequence number to provide an indication of a number of pending messages transmitted to the target;
• and
• upon determining that all messages have been transmitted to the target, removing the tracker entry associated with the initiator from the transmit message queue.