Tracking pipelined activity during off-core memory accesses to evaluate the impact of processor core frequency changes

Patent metadata

Field	Value
Publication number	US-9652356-B2
Application number	US-201514723989-A
Country	US
Kind code	B2
Filing date	May 28, 2015
Priority date	Aug 11, 2014
Publication date	May 16, 2017
Grant date	May 16, 2017

Abstract

A processor system tracks, in at least one counter, a number of cycles in which at least one execution unit of at least one processor core is idle and at least one thread of the at least one processor core is waiting on at least one off-core memory access during run-time of the at least one processor core during an interval comprising multiple cycles. The processor system evaluates an expected performance impact of a frequency change within the at least one processor core based on the current run-time conditions for executing at least one operation tracked in the at least one counter during the interval.

First claim

What is claimed is:

1. A method for tracking pipelined activity during off-core memory accesses, comprising:
tracking, in at least one counter, a number of cycles in which at least one execution unit of at least one processor core is idle and at least one thread of the at least one processor core is waiting on at least one off-core memory access during run-time of the at least one processor core during an interval comprising a plurality of cycles by tracking the number of cycles on at least one multi-cycle execution unit from among the at least one execution unit that requires more than one cycle to complete execution of at least one operation; and
evaluating an expected performance impact of a frequency change within the at least one processor core based on the current run-time conditions tracked in the at least one counter during the interval.

2. The method according to claim 1 , wherein tracking, in at least one counter, a number of cycles in which at least one execution unit of at least one processor core is idle and at least one thread of the at least one processor core is waiting on at least one off-core memory access during run-time of the at least one processor core during an interval comprising a plurality of cycles further comprises:
tracking in a first counter of the at least one counter, each cycle of the plurality of cycles during which none of the at least one execution unit is finishing execution and the at least one thread of the at least one processor core is waiting on at least one off-core memory access; and
tracking in a second counter of the at least one counter, each cycle of the plurality of cycles during which none of the at least one execution unit is busy and the at least one thread of the at least one processor core is waiting on at least one off-core memory access.

3. The method according to claim 2 , further comprising:
evaluating an amount of pipelining overlap within the at least one processor core from the difference between a first value in the first counter and a second value in the second counter.

4. The method according to claim 2 , further comprising:
determining the number of cycles during which the at least one execution unit in the processor core is busy, but none of the plurality of execution units are finishing execution, from a difference between a first value in the first counter and a second value in the second counter.

5. The method according to claim 2 , further comprising:
evaluating the value in the first counter as a measure for an upper bound on a first number of cycles that can be reduced by a frequency change comprising a slower frequency that will not impact performance as measured by a pipeline throughput of the at least one processor core; and
evaluating the value in the second counter as a measure for a lower bound on a second number of cycles that can be reduced by a frequency change comprising a slower frequency that will not impact the performance as measured by the pipeline throughput of the at least one processor core.

6. The method according to claim 2 , further comprising:
tracking, during a first interval running at a first frequency, in the second counter of the at least one counter, each cycle of the plurality of cycles of a workload during which none of the at least one execution unit is busy and the at least one thread of the at least one processor core is waiting on at least one off-core memory access, as a first counter value;
storing the first counter value;
tracking, during a second interval running at a second frequency, in the second counter of the at least one counter, each cycle of the plurality of cycles of the workload during which none of the at least one execution unit is busy and the at least one thread of the at least one processor core is waiting on at least one off-core memory access, as a second counter value, wherein the second frequency is slower than the first frequency;
calculating a separate ratio for each interval, wherein the separate ratio comprises a value in the second counter for the interval over a frequency for the interval; and
determining a change in frequency sensitivity from the first frequency to the second frequency from a difference between a first ratio calculated for the first frequency and a second ratio calculated for the second frequency, wherein the change in frequency sensitivity provides an indicator of the expected performance impact of an additional frequency change within the at least one processor core.

7. The method according to claim 6 , further comprising:
in response to determining that there is not a difference between first ratio and the second ratio, evaluating the lack of difference as an indicator of the extent of dependency among a plurality of execution units of at least one processor core to execute the workload, wherein each of the plurality of execution units is a multi-cycle execution units that requires more than one cycle to complete execution of at least one operation.

8. The method according to claim 6 , further comprising:
in response to determining that there is not a difference between the first ratio and the second ratio, determining an indicator of a change in frequency sensitivity of the workload based on an additional difference between the first counter and the second counter.

9. The method according to claim 2 , further comprising:
tracking, during a first interval running at a first frequency, in the first counter of the at least one counter, each cycle of the plurality of cycles during which none of the at least one execution unit is finishing execution and the at least one thread of the at least one processor core is waiting on at least one off-core memory access, as a first counter value;
storing the first counter value;
tracking, during a second interval running at a second frequency, in the first counter of the at least one counter, each cycle of the plurality of cycles during which none of the at least one execution unit is finishing execution and the at least one thread of the at least one processor core is waiting on at least one off-core memory access, as a second counter value, wherein the second frequency is slower than the first frequency;
calculating a separate ratio for each interval, wherein the separate ratio comprises a value in the first counter for the interval over a frequency for the interval; and
determining a change in frequency sensitivity from the first frequency to the second frequency from a difference between a first ratio calculated for the first frequency and a second ratio calculated for the second frequency, wherein the change in frequency sensitivity provides an indicator of the expected performance impact of an additional frequency change within the at least one processor core.

10. The method according to claim 1 , wherein tracking, in at least one counter, a number of cycles in which at least one execution unit of at least one processor core is idle and at least one thread of the at least one processor core is waiting on at least one off-core memory access during run-time of the at least one processor core during an interval comprising a plurality of cycles:
tracking the number of cycles on at least one multi-cycle execution unit from among the at least one execution unit that requires more than one cycle to complete execution of at least one operation and on at least one single cycle execution unit from among the at least one execution unit that requires one cycle to complete execution of the at least one operation.

11. The method according to claim 1 , wherein tracking, in at least one counter, a number of cycles in which at least one execution unit of at least one processor core are idle and at l

Assignees

• IBM

Inventors

• Hanson Heather L
• Indukuru Venkat R
• O'Connell Francis P
• Rajamani Karthick

Classifications

• Y02B60/1217

  Cross-Sectional Technologies · mapped topic

• G06F1/3206

  Monitoring of events, devices or parameters that trigger a change in power modality · CPC title

• G06F11/3466 Primary

  Performance evaluation by tracing or monitoring · CPC title

• Y02B60/1239

  Cross-Sectional Technologies · mapped topic

• G06F1/3243

  Power saving in microcontroller unit · CPC title