Forwarding element data plane performing floating point computations
Some embodiments provide a network forwarding element with a data-plane forwarding circuit that has a parameter collecting circuit to store and distribute parameter values computed by several machines in a network. In some embodiments, the machines perform distributed computing operations, and the parameter values that compute are parameter values associated with the distributed computing operations. The parameter collecting circuit of the data-plane forwarding circuit (data plane) in some embodiments (1) stores a set of parameter values computed and sent by a first set of machines, and (2) distributes the collected parameter values to a second set of machines once it has collected the set of parameter values from all the machines in the first set. The first and second sets of machines are the same set of machines in some embodiments, while they are different sets of machines (e.g., one set has at least one machine that is not in the other set) in other embodiments. In some embodiments, the parameter collecting circuit performs computations on the parameter values that it collects and distributes the result of the computations once it has processed all the parameter values distributed by the first set of machines. The computations are aggregating operations (e.g., adding, averaging, etc.) that combine corresponding subset of parameter values distributed by the first set of machines.
1. An apparatus comprising:
packet processing circuitry to:
perform computations based on floating-point parameter values extracted from one or more received packets; and
provide results of the computations in one or more packets for transmission to at least one destination, wherein the perform computations or provide results comprise performance of one or more match action operations;
wherein:
the results of the computations are to be provided as one or more key-value pairs in one or more protocol headers of the one or more packets for the transmission to the at least one destination; and
the results of the computations are for use in one or more distributed machine-learning operations.
2. The apparatus of claim 1 , wherein the packet processing circuitry comprises:
at least one programmable packet processing stage to perform packet
forwarding
operations on received packets and
at least one programmable packet processing stage to perform computations and provide results.
3. The apparatus of claim 1 , wherein the packet processing circuitry comprises
at least one programmable packet processing stage that is programmed to perform the computations.
4. The apparatus of claim 1 , wherein the computations comprise one or more of: addition or averaging.
5. The apparatus of claim 1 , wherein the computations are associated with at least one machine learning training operation.
6. The apparatus of claim 2 , wherein the at least one machine learning training operation comprises an all reduce, or all gather operation.
7. The apparatus of claim 1 , wherein the apparatus further comprises a switch that comprises at least one ingress port to receive one or more packets.
8. The apparatus of claim 1 , wherein the apparatus further comprises a switch that comprises at least one egress port to receive one or more packets.
9. A non-transitory computer readable medium storing instructions, that when executed, configure a packet processing circuitry of a forwarding element to perform operations comprising:
perform computations based on floating-point parameter values in one or more received packets received by the forwarding element; and
provide results of the computations in one or more packets for transmission by the forwarding element to at least one destination, wherein the perform computations or provide results comprise performance of one or more match action operations by the packet processing circuitry;
wherein:
the results of the computations are to be provided as one or more key-value pairs in one or more protocol headers of the one or more packets for the transmission to the at least one destination; and
the results of the computations are for use in one or more distributed machine-learning operations.
10. The non-transitory computer readable medium of claim 9 , wherein the packet processing circuitry comprises:
at least one programmable packet processing stage to perform packet forwarding operations on received packets and
at least one programmable packet processing stage to perform the computations and provide results.
11. The non-transitory computer readable medium of claim 9 , wherein the packet processing circuitry comprises:
at least one programmable packet processing stage that is programmed to perform the computations.
12. The non-transitory computer readable medium of claim 9 , wherein the computations comprise one or more of: addition or averaging.
13. The non-transitory computer readable medium of claim 9 , wherein the computations are associated with at least one machine learning training operation.
14. The non-transitory computer readable medium of claim 9 , wherein the at least one machine learning training operation comprises an all reduce or all gather operation.
15. A method comprising:
in a packet processing circuitry of a forwarding element:
perform computations based on floating-point parameter values in one or more received packets received by the forwarding element; and
provide results of the computations in one or more packets for transmission by the forwarding element to at least one destination, wherein the perform computations or provide results comprise performance of one or more match action operations by the packet processing circuitry;
wherein:
the results of the computations are to be provided as one or more key-value pairs in one or more protocol headers of the one or more packets for the transmission to the at least one destination; and
the results of the computations are for use in one or more distributed machine-learning operations.
16. The method of claim 15 , comprising:
in at least one programmable packet processing stage of the packet processing circuitry, performing packet forwarding operations on received packets and
in at least one programmable packet processing stage of the packet processing circuitry, performing the computations and providing the results.
17. The method of claim 15 , wherein the computations comprise one or more of: addition or averaging.
18. The method of claim 15 , wherein the computations are associated with at least one machine learning training operation.
19. The method of claim 18 , wherein the at least one machine learning training operation comprises an all reduce or all gather operation.
20. The method of claim 15 , comprising:
receiving a configuration to program the packet processing circuitry of the forwarding element to perform the computations and provide the results of the computations in the one or more packets for the transmission by the forwarding element to the at least one destination.