may predict and store operation execution times of the NPU 830 and the CPU 840 for processing the third subgraph 530 at the available frequencies. In addition, the profiling unit 120 may predict and store operation execution times of the CPU 850 processor for processing the sixth subgraph 560 at available frequencies.

Examples, in which frequencies available in the CPU 850 processor are about 2.42 GHz, about 1.92 GHz, and about 1.40 GHz, will be examined. The profiling unit 120 may predict and store an operation execution time of the subgraph 560 at about 2.42 GHz of the CPU 850 processor as about 60 ms (852), an operation execution time of the subgraph 560 at about 1.92 GHz as about 90 ms (854), and an operation execution time of the subgraph 560 at about 1.40 GHz as about 120 ms (856).

According to an embodiment, the scheduler 130 may select a subgraph and a processor based on a request received from at least one application and at least one operation execution time stored for each subgraph in the profiling unit 120.

Referring to FIG. 7, the scheduler 130 may include a multi-level queue 132 and a scheduling policy unit 134.

The multi-level queue 132 may include a plurality of queues for satisfying requests of various applications. Each queue may be assigned with a preset policy. In addition, priorities between policies assigned to each of the plurality of queues included in the multi-level queue 132 may be pre-defined.

The scheduler 130 may receive units S122, that are obtained by the analysis unit 110 dividing a request S200 received from at least one application, and input the unis S122 to the multi-level queue 132 with respect to types of units. The multi-level queue 132 may insert the units S122, which are obtained by partitioning the received request S200, into queues having policies corresponding to policies of the received request S200.

For example, it is assumed that the first application App1 101 receives a request with a latency deadline, and the second application App2 102 receives a best-effort request to be performed as quickly as possible. In a controller (e.g., a processor) of the terminal 100, priorities between the latency deadline policy and the best-effort policy are assigned in advance. For example, in the controller, the latency deadline policy may be given a first priority, and the best-effort policy may be given a second priority in advance. The deadline policy is a scheduling policy that considers application service-level objective.

In this case, the multi-level queue 132 may assign a request of the first application App1 101 to a first queue, and assign a request of the second application App2 102 to a second queue. It is assumed to be previously set that the first queue has a latency deadline policy, and the second queue has a best-effort policy.

The scheduling policy unit 134 may sequentially circulate the queues of the multi-level queue 132 according to a preset priority for each policy, and select a subgraph and a processor suitable for processing the pending units in the queue.

A function of the scheduler 130 will be described with reference to FIGS. 6 through 8.

According to embodiments, FIG. 6 and FIG. 10 illustrate an example of selecting a subgraph in the scheduler 130 when all processors are available. FIG. 7 illustrates an example of a method of operating the scheduler 130, according to an embodiment. FIG. 8 illustrates an example of transceiving an operation execution time between the profiling unit 120 and the scheduler 130, according to an embodiment.

FIG. 6 illustrates a case in which the analysis unit 110 receives a first request 610 from the first application App1 101, and receives a second request 620 from the second application App2 102.

The first request 610 may be partitioned into a plurality of units (e.g., unit A 611, unit B 613, and unit C 615). It is assumed that the second request 620 is partitioned into a plurality of units (e.g., unit D 621, unit E 623, and unit F 625), and the unit D 621 has already been processed. In addition, it is assumed that the first request 610 and the second request 620 are requests having a latency deadline policy.

The scheduler 130 may insert the unit A 611, the unit B 613, and the unit C 615 into a queue corresponding to the policy of the first request 610 via the multi-level queue 132. In addition, the scheduler 130 may insert the unit D 621, the unit E 623, and the unit F 625 into a queue corresponding to the policy of the second request 620. In this case, the multi-level queue 132 may insert the first request 610 and the second request 620 into queues having latency deadlines as a policy, respectively. When the first request 610 and the second request 620 have the same policy, the multi-level queue 132 may insert the first request 610 and the second request 620 into the same queue.

The scheduling policy unit 134 may select a subgraph and a processor suitable for processing a pending unit in a queue based on the latency deadline policy. The scheduling policy unit 134 may receive current frequency information from at least one processor (e.g., 180, 182, 184, and 186 in FIG. 1) used in the terminal 100, transmit current frequency information received from the processor to the profiling unit 120, and request an operation execution time for a subgraph generated based on the plurality of units constituting the first request 610 and the second request 620.

For example, referring to FIGS. 6 and 10, when the first request 610 may be processed by using a subgraph 530 including the unit A 611 and the unit B 611, and a subgraph 540 including the unit C 614, the scheduling policy unit 134 may request an operation execution time corresponding to a current frequency (i.e., frequency at the time of receiving the request from the application) among available frequencies of the at least one processor capable of processing each of the subgraphs 530 and 540.

In the embodiment of FIG. 8, the profiling unit 120 may receive a request for the operation execution time S300, may transmit an operation execution time of about 4.33 ms 832 corresponding to a current frequency of about 585 MHz of the NPU 830 rather than operation execution times 832 and 834 of the NPU 830 processor for processing the subgraph 530 including the unit A 611 and the unit B 613 at other available frequencies, and may transmit an operation execution time of about 5.67 ms 822 S310 corresponding to a current frequency of about 2.42 GHz of a CPU 820