Granted Patent B2

US 12,737,230 · App. 18/765,445 · Granted Sep 15, 2026

Inventors: Diman Zad Tootaghaj (Milpitas, CA); Yunming Xiao (Ypsilanti, MI); Aditya Dhakal (Santa Clarita, CA); Puneet Sharma (Palo Alto, CA); Lianjie Cao (Milpitas, CA)

Assignee: Hewlett Packard Enterprise Development LP

G06F9/5038G06F9/5016G06F15/17331

Abstract

Claims (50)

1 . An adapter for a system comprising a plurality of physical parallel processing units (PPUs), the adapter comprising:

a network interface to communicate over a network; and

an adapter controller to:

receive, over the network, job data for multiple jobs to be executed on PPU fractions of one or more physical PPUs,

determine that first job data of a first job of the multiple jobs is to be provided to a first PPU fraction of the PPU fractions, and that second job data of a second job of the multiple jobs is to be provided to a second PPU fraction of the PPU fractions, and

initiate a direct memory access (DMA) transfer of the first job data to a first PPU memory buffer of the first PPU fraction, and a DMA transfer of the second job data to a second PPU memory buffer of the second PPU fraction.

2 . The adapter of claim 1 , comprising a job scheduler executable by the adapter controller to allocate the multiple jobs to the PPU fractions, wherein data isolation is provided between the first job allocated to the first PPU fraction and the second job allocated to the second PPU fraction.

3 . The adapter of claim 2 , wherein the job scheduler is executable by the adapter controller to allocate the multiple jobs to the PPU fractions based on one or more objectives, the one or more objectives selected from among an operational cost reduction objective to reduce a cost associated with a usage of the one or more physical PPUs, or a migration cost reduction objective to reduce a cost associated with migrating jobs between physical PPUs.

4 . The adapter of claim 1 , further comprising:

an adapter memory,

wherein the adapter controller is to provide adapter memory buffers in the adapter memory for respective clients that submitted the multiple jobs, and

wherein a first adapter memory buffer of the adapter memory buffers is to receive the first job data of the first job from a first client, and a second adapter memory buffer of the adapter memory buffers is to receive the second job data of the second job from a second client.

5 . The adapter of claim 4 , wherein the first job data is received in a Remote Direct Memory Access (RDMA) transfer from the first client to the first adapter memory buffer, and the second job data is received in a RDMA transfer from the second client to the second adapter memory buffer.

6 . The adapter of claim 1 , wherein the adapter is separate from a central processing unit (CPU) of the system, and the adapter controller is to:

responsive to a completion of the DMA transfer of the first job data of the first job to the first PPU memory buffer of the first PPU fraction, notify the CPU of the completion to cause invocation of machine-readable instructions by the CPU at the one or more physical PPUs to process the first job data.

7 . The adapter of claim 6 , wherein the adapter controller is to:

retrieve, at the adapter from the CPU, a result of the processing of the first job data, the result retrieved using a DMA transfer from the first PPU memory buffer of the first PPU fraction.

8 . The adapter of claim 7 , wherein the adapter controller is to:

receive an indication from a physical PPU of the one or more physical PPUs or from the CPU that the result is available at the first PPU memory buffer; and

initiate the DMA transfer from the first PPU memory buffer in response to the indication.

9 . The adapter of claim 1 , wherein the adapter controller is to:

receive a first memory address of the first PPU memory buffer of the first PPU fraction reserved by a central processing unit (CPU) of the system, and

receive a second memory address of the second PPU memory buffer of the second PPU fraction reserved by the CPU.

10 . The adapter of claim 9 , wherein the adapter controller is to:

store the first memory address and the second memory address in respective entries of mapping information that contain buffer identifiers of respective PPU memory buffers and memory addresses of the respective PPU memory buffers, the mapping information tracking allocations of PPU memory buffers in the one or more physical PPUs, and

responsive to receiving the first job data of the first job, perform a lookup of the mapping information to obtain the first memory address for accessing the first PPU memory buffer, wherein the DMA transfer of the first job data to the first PPU memory buffer uses the first memory address obtained from the mapping information.

11 . The adapter of claim 10 , wherein each respective entry of the entries of the mapping information further comprises a mapping value derived by applying a function on a memory address and a buffer identifier of a PPU memory buffer, wherein the first job data is associated with metadata comprising a first mapping value, and wherein the lookup of the mapping information uses the first mapping value to retrieve an entry of the mapping information.

12 . The adapter of claim 1 , comprising a smart network interface controller (NIC) including the network interface and the adapter controller.

13 . The adapter of claim 1 , wherein the adapter controller is to:

receive, over the network, the first job data of the first job according to a first format, and

translate the first job data according to the first format to converted job data according to a second format different from the first format,

wherein the DMA transfer of the first job data to the first PPU memory buffer comprises a DMA transfer of the converted job data to the first PPU memory buffer.

14 . The adapter of claim 13 , wherein the first format is a serial format, and the first job data comprises a serial stream of data, and

wherein the translating comprises deserializing the first job data into the converted job data according to the second format.

15 . The adapter of claim 1 , wherein the adapter is separate from a central processing unit (CPU) of the system, and the CPU is not in a data path of the DMA transfer of the first job data to the first PPU memory buffer of the first PPU fraction, and the DMA transfer of the second job data to the second PPU memory buffer of the second PPU fraction.

16 . The adapter of claim 1 , comprising one or more of logic to compress and decompress data associated with jobs executed by the one or more physical PPUs, or logic to encrypt or decrypt data associated with the jobs.

17 . A non-transitory machine-readable storage medium storing instructions that upon execution cause a host central processing unit (CPU) of a system to:

allocate parallel processing unit (PPU) memory buffers in respective PPU fractions of one or more physical PPUs;

send, to an adapter, references to the PPU memory buffers for association in mapping information to jobs from clients, wherein the adapter is separate from the host CPU;

receive, from the adapter, an indication of a direct memory access (DMA) transfer of job data of a job from the adapter to a first PPU memory buffer of a first PPU fraction of the PPU fractions; and

based on the indication, invoke machine-readable instructions in the first PPU fraction to process the job data in the first PPU memory buffer.

18 . The non-transitory machine-readable storage medium of claim 17 , wherein the mapping information is populated with the references to the PPU memory buffers and identifiers indicating the clients based on job scheduling of jobs to the PPU fractions by a job scheduler executed by the adapter.

19 . A method comprising:

receiving, by an adapter over a network:

first job data of a first job transferred from a first client in a first remote direct memory access (RDMA) transfer to a first adapter memory buffer in an adapter memory of the adapter, and

second job data of a second job transferred from a second client in a second RDMA transfer to a second adapter memory buffer in the adapter memory;

determining, by the adapter, that the first job data is to be provided to a first parallel processing unit (PPU) fraction of one or more physical PPUs, and that the second job data is to be provided to a second PPU fraction of the one or more physical PPUs;

performing a direct memory access (DMA) transfer of the first job data from the adapter to a first PPU memory buffer of the first PPU fraction, and a DMA transfer of the second job data from the adapter to a second PPU memory buffer of the second PPU fraction; and

receiving, by the adapter, a first result of processing of the first job data by a first compute resource in the first PPU fraction, and a second result of processing of the second job data by a second compute resource in the second PPU fraction.

20 . The method of claim 19 , wherein the determining that the first job data is to be provided to the first PPU fraction and that the second job data is to be provided to the second PPU fraction is based on looking up a hash map at the adapter, the hash map comprising a plurality of entries that track PPU memory buffers to jobs from clients.

Assignments (1)

Continuity (1)

References Cited (110)