Computational data storage systems
In one embodiment, a system comprises a host processor and a storage system. The storage system comprises one or more storage devices, and each storage device comprises a non-volatile memory and a compute offload controller. The non-volatile memory stores data, and the compute offload controller performs compute tasks on the data based on compute offload commands from the host processor.
1. A storage device, comprising:
a non-volatile storage medium; and
processing circuitry to:
receive, via interface circuitry, a chained compute offload command from a host computing device, wherein the chained compute offload command instructs the storage device to perform a plurality of chained compute operations on an input data object stored on the non-volatile storage medium, wherein the chained compute operations include different types of compute operations;
read the input data object from the non-volatile storage medium;
perform the plurality of chained compute operations on the input data object, wherein an output data object is computed based on performing the plurality of chained compute operations; and
return the output data object as an output of the chained compute offload command.
2. The storage device of claim 1 , wherein the chained compute offload command comprises:
an input parameter indicating a first location of the input data object on the non-volatile storage medium;
an operation chain parameter indicating the chained compute operations to be performed on the input data object and an order in which the chained compute operations are to be performed; and
an operation arguments parameter indicating one or more arguments for the respective chained compute operations.
3. The storage device of claim 2 , wherein the chained compute offload command further comprises:
an output parameter indicating how the output data object is to be returned.
4. The storage device of claim 3 , wherein:
the output parameter indicates that the output data object is to be written to the non-volatile storage medium; and
the processing circuitry to return the output data object as the output of the chained compute offload command is further to:
write the output data object to a second location on the non-volatile storage medium; and
send, via the interface circuitry, an address of the second location to the host computing device.
5. The storage device of claim 4 , wherein the second location comprises a pre-allocated output buffer on the non-volatile storage medium.
6. The storage device of claim 3 , wherein:
the output parameter indicates that the output data object is to be written to a second location on the non-volatile storage medium; and
the processing circuitry to return the output data object as the output of the chained compute offload command is further to:
write the output data object to the second location on the non-volatile storage medium.
7. The storage device of claim 3 , wherein:
the output parameter indicates that the output data object is to be sent to the host computing device; and
the processing circuitry to return the output data object as the output of the chained compute offload command is further to:
send, via the interface circuitry, the output data object to the host computing device.
8. The storage device of claim 2 , wherein the processing circuitry to perform the plurality of chained compute operations on the input data object is further to:
perform the plurality of chained compute operations based on the order indicated by the operation chain parameter and the one or more arguments indicated by the operation arguments parameter;
compute a plurality of intermediate outputs of the plurality of chained compute operations, wherein each intermediate output of a particular chained compute operation is input to a corresponding subsequent chained compute operation based on the order indicated by the operation chain parameter; and
compute the output data object based on the plurality of intermediate outputs.
9. The storage device of claim 8 , wherein the processing circuitry to compute the plurality of intermediate outputs of the plurality of chained compute operations is further to:
write at least a subset of the plurality of intermediate outputs to a buffer pool on the non-volatile storage medium; and
return an address of each intermediate output in the buffer pool to the corresponding subsequent chained compute operation.
10. The storage device of claim 1 , wherein:
the non-volatile storage medium comprises a non-volatile memory; and
the storage device is a block storage device, wherein the block storage device is to provide block-based storage on the non-volatile memory.
11. At least one non-transitory machine-readable medium having instructions stored thereon, wherein the instructions, when executed on processing circuitry of a storage device, cause the processing circuitry to:
receive, via interface circuitry, a chained compute offload command from a host computing device, wherein the chained compute offload command instructs the storage device to perform a plurality of chained compute operations on input data stored on a non-volatile storage medium of the storage device, wherein the chained compute operations include different types of compute operations;
read the input data from the non-volatile storage medium;
perform the plurality of chained compute operations on the input data, wherein output data is computed based on performing the plurality of chained compute operations; and
return the output data as an output of the chained compute offload command.
12. The at least one non-transitory machine-readable medium of claim 11 , wherein the chained compute offload command comprises:
an input parameter indicating a first location of the input data on the non-volatile storage medium;
an operation chain parameter indicating the chained compute operations to be performed on the input data and an order in which the chained compute operations are to be performed; and
an operation arguments parameter indicating one or more arguments for the respective chained compute operations.
13. The at least one non-transitory machine-readable medium of claim 12 , wherein the chained compute offload command further comprises:
an output parameter indicating how the output data is to be returned.
14. The at least one non-transitory machine-readable medium of claim 13 , wherein:
the output parameter indicates that the output data is to be written to the non-volatile storage medium; and
the instructions that cause the processing circuitry to return the output data as the output of the chained compute offload command further cause the processing circuitry to:
write the output data to a second location on the non-volatile storage medium; and
send, via the interface circuitry, an address of the second location to the host computing device.
15. The at least one non-transitory machine-readable medium of claim 14 , wherein the second location comprises a pre-allocated output buffer on the non-volatile storage medium.
16. The at least one non-transitory machine-readable medium of claim 13 , wherein:
the output parameter indicates that the output data is to be written to a second location on the non-volatile storage medium; and
the instructions that cause the processing circuitry to return the output data as the output of the chained compute offload command further cause the processing circuitry to:
write the output data to the second location on the non-volatile storage medium.
17. The at least one non-transitory machine-readable medium of claim 13 , wherein:
the output parameter indicates that the output data is to be sent to the host computing device; and
the instructions that cause the processing circuitry to return the output data as the output of the chained compute offload command further cause the processing circuitry to:
send, via the interface circuitry, the output data to the host computing device.
18. The at least one non-transitory machine-readable medium of claim 12 , wherein the instructions that cause the processing circuitry to perform the plurality of chained compute operations on the input data further cause the processing circuitry to:
perform the plurality of chained compute operations based on the order indicated by the operation chain parameter and the one or more arguments indicated by the operation arguments parameter;
compute a plurality of intermediate outputs of the plurality of chained compute operations, wherein each intermediate output of a particular chained compute operation is input to a corresponding subsequent chained compute operation based on the order indicated by the operation chain parameter; and
compute the output data based on the plurality of intermediate outputs.
19. The at least one non-transitory machine-readable medium of claim 18 , wherein the instructions that cause the processing circuitry to compute the plurality of intermediate outputs of the plurality of chained compute operations further cause the processing circuitry to:
write at least a subset of the plurality of intermediate outputs to a buffer pool on the non-volatile storage medium; and
return an address of each intermediate output in the buffer pool to the corresponding subsequent chained compute operation.
20. A method performed by processing circuitry of a storage device, the method comprising:
receiving a chained compute offload command from a host computing device, wherein the chained compute offload command instructs the storage device to perform a plurality of chained compute operations on an input data object stored on a non-volatile storage medium of the storage device, wherein the chained compute operations include different types of compute operations;
reading the input data object from the non-volatile storage medium;
performing the plurality of chained compute operations on the input data object, wherein an output data object is computed based on performing the plurality of chained compute operations; and
returning the output data object as an output of the chained compute offload command.
21. The method of claim 20 , wherein the chained compute offload command comprises:
an input parameter indicating a first location of the input data object on the non-volatile storage medium;
an operation chain parameter indicating the chained compute operations to be performed on the input data object and an order in which the chained compute operations are to be performed; and
an operation arguments parameter indicating one or more arguments for the respective chained compute operations.
22. The method of claim 21 , wherein performing the plurality of chained compute operations on the input data object comprises:
performing the plurality of chained compute operations based on the order indicated by the operation chain parameter and the one or more arguments indicated by the operation arguments parameter;
computing a plurality of intermediate outputs of the plurality of chained compute operations, wherein each intermediate output of a particular chained compute operation is input to a corresponding subsequent chained compute operation based on the order indicated by the operation chain parameter; and
computing the output data object based on the plurality of intermediate outputs.
23. A system, comprising:
interface circuitry to communicatively couple a host processor and a storage device;
the host processor, wherein the host processor is to send, via the interface circuitry, a chained compute offload command to the storage device, wherein the chained compute offload command instructs the storage device to perform a plurality of chained compute operations on an input data object stored on the storage device, wherein the chained compute operations include different types of compute operations; and
the storage device, wherein the storage device comprises a non-volatile storage medium and compute offload circuitry, wherein the compute offload circuitry is to:
receive, via the interface circuitry, the chained compute offload command from the host processor;
read the input data object from the non-volatile storage medium;
perform the plurality of chained compute operations on the input data object, wherein an output data object is computed based on performing the plurality of chained compute operations; and
return the output data object as an output of the chained compute offload command.
24. The system of claim 23 , wherein the chained compute offload command comprises:
an input parameter indicating a first location of the input data object on the non-volatile storage medium;
an operation chain parameter indicating the chained compute operations to be performed on the input data object and an order in which the chained compute operations are to be performed; and
an operation arguments parameter indicating one or more arguments for the respective chained compute operations.
25. The system of claim 24 , wherein the compute offload circuitry to perform the plurality of chained compute operations on the input data object is further to:
perform the plurality of chained compute operations based on the order indicated by the operation chain parameter and the one or more arguments indicated by the operation arguments parameter;
compute a plurality of intermediate outputs of the plurality of chained compute operations, wherein each intermediate output of a particular chained compute operation is input to a corresponding subsequent chained compute operation based on the order indicated by the operation chain parameter; and
compute the output data object based on the plurality of intermediate outputs.
26. The at least one non-transitory machine-readable medium of claim 11 , wherein:
the input data comprises an input data object; and
the output data comprises an output data object.
27. The storage device of claim 1 , wherein the chained compute offload command comprises one or more parameters, wherein the one or more parameters indicate:
a location of the input data object on the non-volatile storage medium;
the chained compute operations to be performed on the input data object; and
an order in which the chained compute operations are to be performed.
28. The storage device of claim 1 , wherein the plurality of chained compute operations are to transform the input data object into the output data object.