Low Latency Communication Via Memory Windows

摘要

A high performance computer system has compute nodes interconnected by an inter-node communication network. The compute nodes include local packetized interconnects. Shared memory between compute nodes is emulated by tunneling packetized interconnect packets between the compute nodes. A memory region on one of the compute nodes is used as a window into a memory region in another compute node. Special registers may be provided to expedite remote operations on small amounts of data.

主权项

1. A method for exchanging data between compute nodes of a computer system comprising:
sending compute node to a receiving compute node by:
placing a local packetized interconnect packet on a local packetized interconnect of the sending compute node;receiving the local packetized interconnect packet at a network interface of the sending compute node;encapsulating the local packetized interconnect packet in an inter-node communication network packet addressed to the receiving compute node;dispatching the inter-node communication network packet to the receiving compute node by way of the inter-node communication network;receiving the inter-node communication network packet at the network interface of the receiving compute node;extracting the local packetized interconnect packet from the inter-node communication network packet;placing the extracted packet onto the local packetized interconnect of the receiving compute node; and, between receiving the local packetized interconnect packet at the network interface of the sending compute node and placing the extracted packet onto the local packetized interconnect of the receiving compute node, performing an address translation on the local packetized interconnect packet, the address translation comprising writing to the local packetized interconnect packet a translated address based upon an offset value and upon a destination address for the local packetized interconnect packet wherein performing the address translation comprises writing into the local packetized interconnect packet an address determined by:
m=l−A+E+OFFSET where OFFSET is the value of the offset; m is the translated address; l is an initial destination address of the local packetized interconnect packet; A is a base address of the range associated at the sending compute node with the receiving compute node and E is a base address of an address range allocated at the receiving compute node for receiving data from the sending compute node.