In-network computing with reduction operation by switch

One or more non-transitory computer-readable media storing instructions executable by one or more computer processors to perform operations for in-network computing, the operations comprising:

• receiving, by a switch coupled to a group of compute blocks and to a target compute block, a first data packet from a first compute block in the group of compute blocks;
• receiving, by the switch, a second data packet from a second compute block in the group of compute blocks;
• storing, by the switch, the first data packet and the second data packet in one or more buffers associated with the switch;
• after storing the first data packet and the second data packet in the one or more buffers, receiving, by the switch, a third data packet from a third compute block in the group of compute blocks;
• determining, by the switch, whether the one or more buffers have sufficient storage to store the third data packet;
• in response to determining that the one or more buffers do not have sufficient storage to store the third data packet, forming, by the switch, a new data packet by aggregating the first data packet and the second data packet stored in the one or more buffers;
• maintaining, by the switch, a collective sequence comprising a number of bits, each of which corresponding to a respective compute block in the group of compute blocks;
• determining, by the switch, whether a collective sequence number in a fourth data packet is greater than a collective sequence number of the collective sequence;
• and
• in response to determining that the collective sequence number in the fourth data packet is greater than the collective sequence number of the collective sequence, disregarding, by the switch, the fourth data packet.

An apparatus, comprising:

• a computer processor for executing computer program instructions;
• and
• a non-transitory computer-readable memory storing computer program instructions executable by the computer processor to perform operations comprising:
• receiving, by a switch coupled to a group of compute blocks and to a target compute block, a first data packet from a first compute block in the group of compute blocks,
• receiving, by the switch, a second data packet from a second compute block in the group of compute blocks,
• storing, by the switch, the first data packet and the second data packet in one or more buffers associated with the switch,
• after storing the first data packet and the second data packet in the one or more buffers, receiving, by the switch, a third data packet from a third compute block in the group of compute blocks,
• determining, by the switch, whether the one or more buffers have sufficient storage to store the third data packet,
• in response to determining that the one or more buffers do not have sufficient storage to store the third data packet, forming, by the switch, a new data packet by aggregating the first data packet and the second data packet stored in the one or more buffers,
• maintaining, by the switch, a collective sequence comprising a number of bits, each of which corresponding to a respective compute block in the group of compute blocks,
• determining, by the switch, whether a collective sequence number in a fourth data packet is greater than a collective sequence number of the collective sequence, and
• in response to determining that the collective sequence number in the fourth data packet is greater than the collective sequence number of the collective sequence, disregarding, by the switch, the fourth data packet.