US 2020/0266995 A1 · MANAGING STATE IN ACCELERATORS

Abstract

Methods and apparatus for managing state in accelerators. An accelerator performs processing operations on a data chunk relating to a job submitted to the accelerator. During or following processing the data chunk, the accelerator generates state information corresponding to its current state and stores the state information or, optionally, the accelerator state information is obtained and stored by privileged software. In connection with continued processing of the current data chunk or a next job and next data chunk, the accelerator accesses previously stored state information identified by the job and validates the state information was generated by itself, another accelerator, or privileged software. Valid state information is then reloaded to restore the state of the accelerator/process state, and processing continues. The chunk processing, accelerator state store, validation, and restore operations are repeated to process subsequent jobs. An accelerator and/or privileged software may use a MAC (Message Authentication Code) algorithm to generate a MAC over a message comprising the accelerator state information. The MAC is then used to validate previously stored state information.

Claims20

Dependent claims

A method comprising:

using a first accelerator to perform processing operations on a data chunk;
storing state information relating to a state of the first accelerator during or following processing the data chunk, the state information that is stored being generated by the first accelerator or by privileged software;
accessing, via the first accelerator or a second accelerator, stored state information;
validating, via the first accelerator or the second accelerator, that the stored state information that is accessed was generated by the first accelerator or by the privileged software;
and
when the stored state information is valid,
loading the state information to the first accelerator or to the second accelerator;
and
beginning processing a next data chunk or continue processing the data chunk with the first accelerator or the second accelerator.

6 more elements

An accelerator apparatus, comprising:

at least one accelerator including a plurality of processing blocks;
at least one of an input-output (TO) interface and a memory controller;
wherein the accelerator apparatus is configured to be implemented in a platform including memory, and wherein to accelerator apparatus is configured to,
perform, using the plurality of processing blocks corresponding to a first accelerator, processing operations on a data chunk;
generate state information relating to a state of the first accelerator during or following processing the data chunk;
store the state information that is generated in the memory;
access previously stored state information from the memory,
validate the previously stored state information was generated by the first accelerator;
and
when the previously stored state information is valid,
reload the previously stored state information to the first accelerator or load the previously stored state information to a second accelerator;
and
begin processing operations on a next data chunk or continue processing operations on the data chunk using the first accelerator or the second accelerator.

10 more elements

A computer platform, comprising:

a processor, including,
a plurality of processor cores,
an interconnect fabric, communicatively coupled to the plurality of processor cores;
a memory controller, communicatively coupled to the interconnect fabric and having one or more memory channels;
a first input-output (TO) interface, communicatively coupled to the interconnect fabric;
system memory comprising one or more memory devices, each communicatively coupled to at least one memory channel of the memory controller;
and
an accelerator device comprising at least one accelerator including a plurality of processing blocks;
a second IO interface coupled to the first IO interface via a link;
wherein the accelerator device is configured to,
perform, using the plurality of processing blocks for a first accelerator, processing operations on a data chunk;
generate state information relating to a state of the first accelerator during or following processing the data chunk;
store the state information that is generated in the system memory;
access previously stored state information from the system memory,
validate the previously stored state information was generated by the first accelerator;
and
when the previously stored state information is valid,
reload the previously stored state information to the first accelerator to restore the state of the accelerator device or load the previously stored state information to a second accelerator;
and
begin processing a next data chunk or continue processing the data chunk with the first accelerator or the second accelerator.

17 more elements

Dependencies are inferred from the claim text.

Managing state in accelerators

Life of this patent

Managing state in accelerators

Life of this patent

Abstract

Claims20